# Dscvry

> Dscvry builds and deploys models, agents and AI systems that automate complex work for industry, government and defence, from cloud to air-gapped.

Dscvry (Discovery AI Limited, UK; ReDscvry Technology Private Limited, India) builds and deploys models, agents, products, APIs and data systems, and deploys third-party models and products inside a client's boundary. Every system is tested on the client's own work before it is relied on, and runs wherever the business requires, from managed cloud to fully air-gapped.

## What we build

### Models: Specialist models, built and tested.
We build our own models, train them for you, and deploy third-party models when they fit. Each is tested on the work it will do, and runs wherever the work is: hosted, public cloud, private cloud, on-premise, edge, offline or air-gapped. The client owns the finished model's weights.
- [Resolve](https://dscvryai.com/technology/models/resolve/): Reads business documents and makes the call: pay or hold, on time or late, who signs off.
- [Match](https://dscvryai.com/technology/models/#match): Decides whether two records are the same company, and matches automatically only when the evidence is there.
- [Extract](https://dscvryai.com/technology/models/#extract): Our own document reading models: complex documents rebuilt as structured data.
- [PII](https://dscvryai.com/technology/models/#pii): Finds personal and sensitive data in documents, so it can be redacted before it leaves your control.
- [NDA clause model](https://dscvryai.com/insights/nda-clause-model/): Rules on 17 standard NDA clauses and quotes its evidence: 84.5% on a sealed test set, running on an ordinary CPU.
- [Custom models](https://dscvryai.com/technology/models/#make): From your data to a working model, for teams that don’t have AI specialists of their own.
- [Fine-tuning](https://dscvryai.com/technology/models/#make): We teach the model your codes, units and terms, so it reads them the way your engineers, lawyers and analysts do.
- [Domain training](https://dscvryai.com/technology/models/#make): We train it on your industry’s documents, and your experts correct it until it judges the way they do.
- [Smaller models](https://dscvryai.com/technology/models/#make): The quality of a large model in a small one that runs cheaply on ordinary hardware.
- [Third-party models](https://dscvryai.com/technology/models/#third-party): Open models, chosen and tested on your own cases, then run privately in your environment.

### Resolve: Introducing Resolve, our fast decision-making model.
Resolve is a decision model. It reads a case file once and scores every option of every question you ask, so your workflow can act on confident answers and send the rest to a person. It does not write prose, and it does not call any outside service while it works, so it can run inside your own environment. Two editions: Resolve Workflow and Resolve Evidence. On EDB, our synthetic enterprise decision benchmark (development set, 536 questions): Resolve Workflow 87.1%, Resolve Evidence 86.2%; on the locked test set (228 case files) 83.1% and 83.6%, level with its base model Kev (83.7%). Private deployment in your environment, or hosted by Dscvry, after an evaluation on your own cases.
- [Resolve Workflow](https://dscvryai.com/technology/models/resolve/results/#resolve-workflow): Trained for everyday enterprise decisions: invoices and expenses, contracts and renewals, IT operations, support tickets. Strongest on routing questions (who approves, who handles); accuracy level with the most accurate models we tested.
- [Resolve Evidence](https://dscvryai.com/technology/models/resolve/results/#resolve-evidence): Trained for messy, conflicting material: which version, approval, override or record actually governs. 84% fewer errors after our training when finding the evidence that decides a case; whole sets of related cases answered correctly 129 times in 150, against 59 before our training; when one deciding fact changes, it follows the change 92% of the time, against 40%.

### Decision models compared: Enterprise decision models compared.
Eight decision models tested on the same enterprise case files: accuracy, automation, risky approvals, long documents, robustness and speed. Every model's results are shown, including where our own model, Resolve, falls short.
- [Summary](https://dscvryai.com/technology/models/resolve/compare/#summary): Resolve, Kev and Perplexity's decider are level at 86–87% on enterprise case files; Cloudflare, Microsoft, AWS and TypeSafe score 76–80%. Approvals that should have been held: Resolve 8–9 of 236, the others 13 to 35.
- [Most accurate](https://dscvryai.com/technology/models/resolve/compare/#q-accuracy): Resolve Workflow, Resolve Evidence, Kev-27B and Perplexity's decider are statistically level at 85.6–87.1% on EDB.
- [Deciding on their own](https://dscvryai.com/technology/models/resolve/compare/#q-auto): With a threshold set from 50 labelled questions per workflow, Resolve, Kev and Perplexity's decider automated 73–76% of decisions at about 6–7% error.
- [Should-hold approvals](https://dscvryai.com/technology/models/resolve/compare/#q-approve): Of 236 approval-type questions: Resolve Evidence 8, Resolve Workflow 9, Clef 13, Kev 14, Strands-Decider 15, Perplexity's decider 20, Decision-1 24, Jev 35.
- [Long documents](https://dscvryai.com/technology/models/resolve/compare/#q-long): At about 56,000 tokens, Kev-based models including Resolve stayed at about 68–70%; Clef fell to 45.1% and Strands-Decider to 42.9%.
- [Speed and cost](https://dscvryai.com/technology/models/resolve/compare/#q-speed): Both Resolve editions, on one H100 GPU, took a median of 0.36–0.37 seconds per case file; general reasoning models took 13 to 58 seconds per case.

### Agents: Work that runs on its own, and stops for a person.
Agents that plan their work, check every step and ask a person before they act. Each agent has only the tools its job needs, consequential steps wait for a named person, and every run is logged so it can be replayed and explained.
- [Checking](https://dscvryai.com/technology/agents/#checking): Invoices against purchase orders, claims against contracts, reports against the rules.
- [Investigation](https://dscvryai.com/technology/agents/#investigation): Traces a break, a complaint or an incident across your systems, and writes up the evidence.
- [Watching](https://dscvryai.com/technology/agents/#watching): Monitors feeds, regulations, contracts and deadlines, and raises only what matters.
- [Follow-through](https://dscvryai.com/technology/agents/#follow-through): Chases missing documents, approvals and replies until the job is done.
- [Groundwork](https://dscvryai.com/technology/agents/#groundwork): Assembles packs, drafts reports and fills in forms from your own records.
- [Routing](https://dscvryai.com/technology/agents/#routing): Routes cases, updates records and hands each one to the right team.
- [Custom agents](https://dscvryai.com/technology/agents/#custom-agents): We design, build and run agents for the work your team does today, inside your boundary.
- [A plan before any action](https://dscvryai.com/technology/agents/#make): The agent writes down its steps first, so a person can see what it intends to do, and which step will need them.
- [Only the tools it’s allowed](https://dscvryai.com/technology/agents/#make): Every tool is typed and whitelisted per agent. Reading is not writing, and some actions are never available at all.
- [A person signs the consequential step](https://dscvryai.com/technology/agents/#make): Postings, payments and closures wait for a named person, who sees the evidence the agent used and can approve, edit or reject.
- [Every step, kept](https://dscvryai.com/technology/agents/#make): Each plan, call, input, result, retry, refusal and approval is logged, so any run can be replayed and explained.

### Products: Ready-made, for the common jobs.
Ready-made products for recurring jobs, from documents and contracts to sustainability data. Each is a reusable workflow built from tested components, used as it is or as the start of a client's own, and run inside the client's boundary.
- [Structurify](https://dscvryai.com/technology/products/structurify/): Documents turned into structured, review-ready data, with confidence scores and human review.
- [SafetyDataSheetAPI](https://safetydatasheetapi.com/): Scanned or digital safety data sheets turned into data your systems can use, across all 16 sections.
- [Clarify](https://dscvryai.com/technology/products/clarify/): Ask your enterprise data questions in everyday language, and get answers within each person’s access, with sources.
- [GRID](https://dscvryai.com/technology/products/grid/): Shows what AI coding tools changed in your software, how each change was checked, and what was missed.
- [Axiom](https://dscvryai.com/solutions/legal/): Turns contracts and legal documents into searchable data, for faster reviews, early warning of risks and audit-ready reports.
- [Sustainize](https://dscvryai.com/solutions/compliance/): Brings your sustainability data together, from raw figures to board-ready reports, with every number traceable.
- [C-Centric](https://dscvryai.com/technology/products/#c-centric): Footfall, sales, incident and tenant signals in one view, for mall operators and large-format retail.

### GRID: The evidence behind AI-assisted software.
GRID connects AI activity, software changes and validation results, and shows what is known about the work and where the record is incomplete: for assurance, cost, productivity and risk decisions.
- [Compliance & assurance](https://dscvryai.com/technology/products/grid/#compliance-assurance): Can the organisation show how AI-assisted software was produced and checked?
- [Cost](https://dscvryai.com/technology/products/grid/#cost): What work is AI spending supporting?
- [Productivity](https://dscvryai.com/technology/products/grid/#productivity): Where is AI helping work progress, and where is effort consumed without a satisfactory result?
- [Operational risk](https://dscvryai.com/technology/products/grid/#operational-risk): Where does AI-assisted work need closer attention?
- [Oversight](https://dscvryai.com/technology/products/grid/#oversight): Where is visibility reliable across the software estate, and where does it break down?

### Structurify: Stop typing documents in.
Structurify turns documents into structured, review-ready data: pre-built templates, confidence scores, human review, and export to CSV, JSON or webhooks.
- [Templates](https://dscvryai.com/technology/products/structurify/#can): Pre-built templates for invoices, tax forms, medical records, contracts and more, or start blank for your own documents.
- [Review](https://dscvryai.com/technology/products/structurify/#can): Extracted data arrives with confidence scores. People approve, edit or flag values before anything is exported.
- [Formats](https://dscvryai.com/technology/products/structurify/#can): PDFs, images, Word, Excel and PowerPoint files, including scans, in batches. Out come CSV, JSON or webhooks.
- [Plain English](https://dscvryai.com/technology/products/structurify/#can): Extractions are defined in plain English, with no code and no machine learning expertise required.

### Clarify: Ask your data. Check the answer.
Clarify (formerly DotPlot) lets people ask questions of enterprise data in everyday language and get answers with their sources, within the access they already have.
- [Ask](https://dscvryai.com/technology/products/clarify/#can): People ask what they need to know in everyday language, and follow up the way they would with a colleague.
- [Permissions](https://dscvryai.com/technology/products/clarify/#can): Governance-aware access controls mean an answer only draws on data the person asking is allowed to see.
- [Sources](https://dscvryai.com/technology/products/clarify/#can): Each answer shows the rows, filters and tables behind it, so it can be checked rather than trusted.
- [Beyond dashboards](https://dscvryai.com/technology/products/clarify/#can): Business users get answers directly, instead of waiting for an analyst to build or change a dashboard.

### APIs: Plug it into the systems you run.
Model, agent and data APIs that plug into the systems a business already runs, with no migration first. Where a system has no API, the AI reads its screens, reports or exports. Read and write permissions are set and logged per system.
- [Model APIs](https://dscvryai.com/technology/apis/#model-apis): Call our model families from your own systems. Every output comes back with its source.
- [Your model, served](https://dscvryai.com/technology/apis/#your-model-served): The models we train for you, served through the same API and run on your own servers.
- [Solution APIs](https://dscvryai.com/technology/apis/#solution-apis): Production loss, reconciliation or contract review, built into the systems your teams already use.
- [Agent APIs](https://dscvryai.com/technology/apis/#agent-apis): Hand a task to an agent and get the result back, including where it stopped for a person.
- [Data APIs](https://dscvryai.com/technology/apis/#data-apis): The data we collect, clean and match for you, served to any system that needs it.
- [Events](https://dscvryai.com/technology/apis/#events): Your systems get an alert the moment something needs attention, with the evidence attached.
- [Custom APIs](https://dscvryai.com/technology/apis/#custom-apis): We wrap your models, data and workflows in an API your teams and partners can call, wherever it runs.
- [Connected, with the permissions you choose](https://dscvryai.com/technology/apis/#make): ERP, maintenance, document, ticketing and plant systems reached through whatever each one offers, with read and write set per system.
- [No API? That’s fine](https://dscvryai.com/technology/apis/#make): Where there is no clean interface, the AI reads the screens, the nightly reports or the exports, and checks them against each other.
- [Where your team already looks](https://dscvryai.com/technology/apis/#make): Findings appear inside the tools people already have open, next to the record they’re working on.
- [Or call it from your own systems](https://dscvryai.com/technology/apis/#make): Your own software can call the AI directly and get back a result with its reasons and sources.

### Data: The data behind every answer.
Data collected, extracted, matched and labelled from where it lives, without migrating systems first. Every value is tied to its source, kept inside the client's boundary and never used to train models for anyone else.
- [Collection](https://dscvryai.com/technology/data/#collection): We gather documents, databases, sensors, field reports, video and audio from wherever they live, without migrating your systems first.
- [Extraction](https://dscvryai.com/technology/data/#extraction): We turn reports, logs, drawings, forms and scans into structured data, each value tied to where it came from.
- [Processing](https://dscvryai.com/technology/data/#processing): We remove duplicates, standardise units and formats, and match the same asset, supplier or customer across systems.
- [Labelling](https://dscvryai.com/technology/data/#labelling): People who know your industry label the examples your AI learns from, and we create realistic extra examples where real ones are rare.
- [Pipelines](https://dscvryai.com/technology/data/#pipelines): We build and run the connections that keep data moving into the systems you use, and fix them when a source changes.
- [Analysis](https://dscvryai.com/technology/data/#analysis): We find the patterns, anomalies and trends that matter, and show the records behind every finding.
- [Reporting](https://dscvryai.com/technology/data/#reporting): Dashboards and reports built for the people who act on them, from the plant floor to the board pack.
- [Governance](https://dscvryai.com/technology/data/#governance): Who can see what, where every number came from, and sensitive details hidden, with every step recorded.

### Evaluations: Know how well it works, before you rely on it.
Evaluations built from the client's own cases: labelled test sets kept apart from training, field-by-field scoring, accuracy at the client's automation rate, and reruns whenever a model, prompt or source changes. Measured on the client's cases, not a public leaderboard; every model named with its licence.
- [Test sets](https://dscvryai.com/technology/evaluations/#test-sets): Real cases from your own work, labelled by people who know it, and kept apart from anything a model learns from.
- [Scoring](https://dscvryai.com/technology/evaluations/#scoring): Every answer checked field by field against the right one, with accuracy reported at the automation rate you choose.
- [Model comparisons](https://dscvryai.com/technology/evaluations/#model-comparisons): Candidate models, ours or third-party, run on the same cases, each named with its licence.
- [Sign-off checks](https://dscvryai.com/technology/evaluations/#sign-off-checks): Requests that need a person are counted on their own, so you see which ones a model would have let through.
- [Reruns on change](https://dscvryai.com/technology/evaluations/#reruns-on-change): The same tests run again whenever a model, prompt or source changes, before the change goes live.
- [Run records](https://dscvryai.com/technology/evaluations/#run-records): Every run kept with its cases, scores and errors, so any result can be traced back to the cases behind it.

## What it can do (capabilities)

### Sense: Sees, hears and reads the work as it happens.
Sense is how the AI sees, watches, reads and listens: video, images, depth, documents and sound, turned into structured signal the rest of the system can work with.
- [Video](https://dscvryai.com/technology/sense/video/): Camera feeds, live or recorded. Tracks, zones, counts and events.
- [Images](https://dscvryai.com/technology/sense/images/): Camera traps, line cameras, drones, site photos. Species, defects, conditions and markings.
- [Depth & 3D](https://dscvryai.com/technology/sense/depth/): Ordinary images and drone surveys. Volumes, distances, clearances and progress.
- [Documents](https://dscvryai.com/technology/sense/documents/): Scans, PDFs, forms, stamps and handwriting. Fields tied to where they sit on the page.
- [Speech & audio](https://dscvryai.com/technology/sense/audio/): Calls, radio, voice notes, machine sounds. Transcripts, key moments and acoustic alerts.

### Video: Cameras that notice.
How the AI watches video: tracking people and vehicles, restricted zones, counting, PPE and unattended objects, with every alert carrying its clip, track and rule.
- [Tracking](https://dscvryai.com/technology/sense/video/#can): Every person, vehicle and object gets a track: where it came in, where it went, how long it stayed. Behaviour is judged over time, not frame by frame.
- [Zones & perimeters](https://dscvryai.com/technology/sense/video/#can): Draw restricted areas, perimeters and one-way routes on the camera view. The AI flags the moment a track crosses in, and shows the route it took to get there.
- [Counting & flow](https://dscvryai.com/technology/sense/video/#can): Doorways and lines become counters: people in and out, occupancy, queues forming. Kept per camera and per hour, ready for operations and reporting.
- [Safety & PPE](https://dscvryai.com/technology/sense/video/#can): Hard hats, high-visibility clothing and exclusion zones around moving machinery, checked continuously instead of once at the start of a shift.

### Images: Every image, understood.
How the AI understands images: species and individuals, injuries and condition, defects, and the text on nameplates and labels, with every label tied to the region it came from.
- [Species & individuals](https://dscvryai.com/technology/sense/images/#can): Species first, then the individual where markings allow it: ear notches, horn shape, stripes and scars matched against your own records.
- [Injury & condition](https://dscvryai.com/technology/sense/images/#can): Posture, gait, wounds and body condition across a burst of frames, so an injured animal is raised the morning it is seen, not the week after.
- [Defects](https://dscvryai.com/technology/sense/images/#can): Cracks, burrs, marks and missing features, measured and judged against your acceptance limits and your own golden samples.
- [Labels & markings](https://dscvryai.com/technology/sense/images/#can): Nameplates, labels, tags and gauges read straight from a site photo, so the record matches the equipment that is actually installed.

### Documents: Read the page like an expert.
How the AI reads documents: layout, tables, forms, handwriting, stamps and signatures, with every value tied to where it sits on the page and checked against ground truth.
- [Layout & tables](https://dscvryai.com/technology/sense/documents/#can): Sections, tables, rows and columns rebuilt from the page, so a composition table in a safety data sheet comes out as a table, with every CAS number checked.
- [Forms & handwriting](https://dscvryai.com/technology/sense/documents/#can): Permits, checklists and inspection forms read field by field, including handwriting, so an empty signature box is noticed rather than skipped.
- [Stamps & signatures](https://dscvryai.com/technology/sense/documents/#can): Stamps, signatures and dates located and checked against who is allowed to sign, and when the signing should have happened.
- [Ground truth](https://dscvryai.com/technology/sense/documents/#can): Every field is scored against documents a person has checked by hand, kept apart from anything used to build the extractor. Fields that don't agree go to a person.

### Depth: Depth from a single camera.
How the AI turns ordinary images into depth: volumes, distances, clearances and progress, with every figure keeping the surface it was measured from.
- [Volume](https://dscvryai.com/technology/sense/depth/#can): Stockpiles, bins and heaps measured from a drone flight or a fixed camera, and compared with what the books say should be there.
- [Distance & separation](https://dscvryai.com/technology/sense/depth/#can): Distances between people and machines measured as work happens, against the separation your procedures require.
- [Clearance](https://dscvryai.com/technology/sense/depth/#can): Load heights and overhead clearances checked at racks, doors and bridges before the vehicle gets there.
- [Progress](https://dscvryai.com/technology/sense/depth/#can): The surface as built compared with the design, section by section: what's been cut, what's been filled and what's left.

### Audio: Hear what was actually said.
How the AI listens: transcribing calls and radio, separating speakers, catching commitments and complaints, and hearing when a machine starts to sound wrong.
- [Speech & speakers](https://dscvryai.com/technology/sense/audio/#can): Calls, meetings and voice notes transcribed and split by speaker, with names, numbers and references matched to your systems.
- [Key moments](https://dscvryai.com/technology/sense/audio/#can): Complaints, promises, cancellations and distress picked out of a conversation, each linked to the second it was said.
- [Machine sounds](https://dscvryai.com/technology/sense/audio/#can): Microphones on pumps, fans and gearboxes compared with how they sounded last month, so a new tone is noticed while there is time to act.
- [Radio & noisy audio](https://dscvryai.com/technology/sense/audio/#can): Site and control-room radio transcribed through the noise, and linked to the permits, equipment and people it mentions.

### Understand: From piles of paper to records you can use.
Understand turns what the AI senses into structured, searchable records, and answers questions with the source behind every fact.
- [Extract](https://dscvryai.com/technology/understand/#can): Fields, tables and line items taken off the page into a schema-ready record, and checked against the schema before anything downstream sees it.
- [Classify](https://dscvryai.com/technology/understand/#can): Emails, documents and requests labelled by type, priority and owner as they arrive, with the reason for each label kept, so the right team starts on the right thing.
- [Structure](https://dscvryai.com/technology/understand/#can): The same fact written three different ways becomes one consistent record: same fields, same units, same dates.
- [Summarize](https://dscvryai.com/technology/understand/#can): Long documents and threads condensed into the decisions, dates and open questions, with each point linked to the passage it came from.
- [Translate](https://dscvryai.com/technology/understand/#can): Supplier emails, contracts and reports handled in the language they arrive in, with part numbers, dates and amounts carried across exactly.
- [Search](https://dscvryai.com/technology/understand/#can): Reports, work orders, manuals and logs searched together, with each result showing where it came from and why it matched.
- [Answer](https://dscvryai.com/technology/understand/#can): Questions answered from your own documents, with a citation for every fact. Where the sources don't say, the AI says so instead of guessing.

### Check: Every line compared. Only the gaps reach you.
Check compares, reconciles, validates, redacts and detects: every record checked against the rules, the contract and every other source, with the evidence attached.
- [Compare](https://dscvryai.com/technology/check/#can): Two versions of a contract, drawing or policy set side by side, with every real change shown in place and changes of wording alone kept apart.
- [Reconcile](https://dscvryai.com/technology/check/#can): Two, three or more sources matched line by line, with each break shown where it happens and why.
- [Match](https://dscvryai.com/technology/check/#can): Suppliers, customers and parts matched across systems by what identifies them, such as tax numbers, bank accounts and addresses, not by how the name is spelled.
- [Validate](https://dscvryai.com/technology/check/#can): Every record checked against your rules and contract terms, not a sample, with the evidence for each pass and fail.
- [Detect](https://dscvryai.com/technology/check/#can): Duplicates, outliers and patterns that sit just under a limit, found across all the records rather than one at a time.
- [Redact](https://dscvryai.com/technology/check/#can): Names, addresses, card numbers and health details found and removed before a record is shared, with the original kept under access control.

### Investigate: Find out what happened, and why.
Investigation across a client's systems: tracing an order, payment, complaint or incident through every system it touched, rebuilding the timeline from the records, naming the likely cause with the evidence for and against, and writing it up with every event linked to its source for a person to check.
- [Trace](https://dscvryai.com/technology/investigate/#trace): Follow one order, payment, complaint or incident through every system it touched.
- [Correlate](https://dscvryai.com/technology/investigate/#correlate): Line up calls, emails, logs, camera feeds and records on one timeline.
- [Reconstruct](https://dscvryai.com/technology/investigate/#reconstruct): Rebuild what happened, step by step, from the records that exist.
- [Explain](https://dscvryai.com/technology/investigate/#explain): Name the most likely cause, and show the evidence for and against it.
- [Write up](https://dscvryai.com/technology/investigate/#write-up): Draft the findings with every event linked to its source, ready for a person to check.

### Anticipate: Know what comes next, and plan for it.
Anticipate turns what already happened into a forward view: failures, shortages, missed deadlines and recurring causes, each with how likely it is and when.
- [Predict](https://dscvryai.com/technology/anticipate/#can): The chance of failure day by day, from the asset's own history and how its past failures began, set against your next planned stop.
- [Forecast](https://dscvryai.com/technology/anticipate/#can): Demand forecast from usage, plans and known events like shutdowns, so orders go in while there is still time.
- [Simulate](https://dscvryai.com/technology/anticipate/#can): A new shift pattern, price or supplier run against your own history first, so you see what it does to delivery, cost and workload before anything changes.
- [Plan](https://dscvryai.com/technology/anticipate/#can): Jobs, people and parts fitted around your limits, such as crews, permits and shutdown windows, and rebuilt when a part runs late or a job overruns.
- [Recommend](https://dscvryai.com/technology/anticipate/#can): When something needs a decision, the choices are laid out in order with what each one costs and when it lands, so your team makes the call faster.

### Operate: The next step, taken. The reason, recorded.
Operate connects the AI to the systems you run, watches for events, routes decisions to the right person and acts within limits you set, with every step logged.
- [Decide](https://dscvryai.com/technology/operate/#can): Each case approved, escalated or held by your own rules, with the rule and the evidence saved next to the decision.
- [Route](https://dscvryai.com/technology/operate/#can): Each decision routed by your rules: who approves what, at which amount, and what needs a second check.
- [Draft](https://dscvryai.com/technology/operate/#can): Replies, reports and forms drafted from your own records and policies, in the tone you use, and sent only when a person signs them off.
- [Converse](https://dscvryai.com/technology/operate/#can): Customers and staff helped by chat or phone with answers from your own records, and passed to a person, with the conversation attached, when a case needs judgement.
- [Act](https://dscvryai.com/technology/operate/#can): The AI drafts, records and raises. What it may never do, it can't, however it is asked.
- [Connect](https://dscvryai.com/technology/operate/#can): ERP, CMMS, CRM, email and document stores connected with the permissions you choose, from read-only upwards.
- [Observe](https://dscvryai.com/technology/operate/#can): New records, signatures, readings and emails watched as they happen; the ones that matter start a workflow.

## Use cases

Illustrative worked examples: the questions operations, government, defence, finance, legal and service teams bring us. Each describes what the AI can do; none claims a client.

### Operations
- [Plant restart](https://dscvryai.com/use-cases/plant-restart/): "What’s blocking the restart?" Decide what’s ready and what’s blocking restart.
- [Equipment failure](https://dscvryai.com/use-cases/equipment-failure/): "What fails next?" Spot the patterns and warn what may fail next.
- [Maintenance planning](https://dscvryai.com/use-cases/maintenance-planning/): "What needs maintenance, and when?" Decide what should be maintained, and when.
- [Stock shortages](https://dscvryai.com/use-cases/stock-shortages/): "What runs out before the next delivery?" Anticipate shortages before they become a problem.
- [Wildlife monitoring](https://dscvryai.com/use-cases/wildlife-monitoring/): "Which animals passed, and are they hurt?" Identify species, spot injuries and flag poachers from camera traps.

### Government
- [Casework backlog](https://dscvryai.com/use-cases/casework-backlog/): "Which cases can be decided today?" Clear the routine cases fast, and send the hard ones to a caseworker.
- [Records release](https://dscvryai.com/use-cases/records-release/): "What must be redacted before release?" Find personal and sensitive details across a release, with every redaction shown for review.
- [Consultation responses](https://dscvryai.com/use-cases/consultation-responses/): "What are the responses actually saying?" Read every response, group the arguments, and quote the evidence for each.
- [Bid evaluation](https://dscvryai.com/use-cases/bid-evaluation/): "Does this bid meet every requirement?" Score each bid against the tender, with every finding linked to the page it came from.
- [Grant claims](https://dscvryai.com/use-cases/grant-claims/): "Which grant claims don’t add up?" Check claims against the rules and the evidence, and flag the ones that need a closer look.

### Defence
- [Overnight reporting](https://dscvryai.com/use-cases/overnight-reporting/): "What changed in the reporting overnight?" Read reports and feeds in many languages, and brief what changed, with sources.
- [Equipment readiness](https://dscvryai.com/use-cases/equipment-readiness/): "Which vehicles are ready to deploy?" See what is fit, what is due, and what is holding the rest back.
- [Site change](https://dscvryai.com/use-cases/site-change/): "What changed at this site since last week?" Compare images over time and flag what is new, moved or missing.
- [Perimeter breaches](https://dscvryai.com/use-cases/perimeter-breaches/): "Is anyone where they shouldn’t be?" Catch people and vehicles crossing the line, day or night.
- [Technical manuals](https://dscvryai.com/use-cases/technical-manuals/): "Where is the fix for this fault?" Answers from your manuals and maintenance records, offline, with the page they came from.

### Supply chain
- [Late shipments](https://dscvryai.com/use-cases/late-shipments/): "Will this shipment arrive on time?" Spot delays early and tell the customer first.
- [Proof of delivery](https://dscvryai.com/use-cases/proof-of-delivery/): "Does the delivery match the order?" Match delivery notes, photos and invoices.
- [Freight audit](https://dscvryai.com/use-cases/freight-audit/): "Are we overpaying on freight?" Check carrier bills against agreed rates.
- [Demand planning](https://dscvryai.com/use-cases/demand-planning/): "How much will we need next month?" Plan stock and capacity ahead of demand.

### Quality & safety
- [Defect inspection](https://dscvryai.com/use-cases/defect-inspection/): "Is this part defective?" Catch defects on the line, not at the customer.
- [Supplier quality](https://dscvryai.com/use-cases/supplier-quality/): "Is the supplier’s evidence complete?" Find missing or inconsistent evidence.
- [Safety checks](https://dscvryai.com/use-cases/safety-checks/): "What here isn’t safe?" Flag what’s unsafe or incomplete.
- [Regulatory changes](https://dscvryai.com/use-cases/regulatory-changes/): "What does this new rule change for us?" Map new rules to what the business must change.

### Finance
- [Invoice errors](https://dscvryai.com/use-cases/invoice-errors/): "Which invoices are we overpaying?" Catch overbilling, duplicates and mistakes before payment.
- [Purchase orders](https://dscvryai.com/use-cases/purchase-orders/): "Does this PO match what we’re paying?" Find mismatches before payment.
- [Revenue leakage](https://dscvryai.com/use-cases/revenue-leakage/): "What should we have charged?" Find money that should have been charged.
- [Onboarding checks](https://dscvryai.com/use-cases/onboarding-checks/): "Is this customer who they say they are?" Verify identity documents and flag the gaps.

### Legal
- [Contract obligations](https://dscvryai.com/use-cases/contract-obligations/): "What does this contract commit us to?" Every obligation, owner and deadline in one list.
- [Contract compliance](https://dscvryai.com/use-cases/contract-compliance/): "Did they deliver what the contract says?" See whether obligations were actually met.
- [Sensitive data](https://dscvryai.com/use-cases/sensitive-data/): "What personal data is in these files?" Find it and remove it before anything is shared.
- [Audit evidence](https://dscvryai.com/use-cases/audit-evidence/): "Where is the evidence for the audit?" Pull the proof together from every system.

### Customer service
- [Call reasons](https://dscvryai.com/use-cases/call-reasons/): "Why are customers calling?" Find what drives calls, straight from the recordings.
- [Customer complaints](https://dscvryai.com/use-cases/customer-complaints/): "What actually happened with this complaint?" Reconstruct what happened, and how it was handled.
- [Claims triage](https://dscvryai.com/use-cases/claims-triage/): "Which claims can we pay today?" Settle the simple ones fast. Send the rest to a person.
- [Customer churn](https://dscvryai.com/use-cases/customer-churn/): "Who is about to leave?" See which accounts are drifting, and why.

### Data
- [Ask your data](https://dscvryai.com/use-cases/ask-your-data/): "How did the North region do last quarter?" Plain-language answers from your own databases, within each person’s access.
- [Data quality](https://dscvryai.com/use-cases/data-quality/): "Can we trust these numbers?" Find duplicates, gaps and records that disagree across systems.
- [Reporting](https://dscvryai.com/use-cases/reporting/): "Why does month-end reporting take a week?" Pull the figures together automatically, every number traced to its source.
- [Data readiness](https://dscvryai.com/use-cases/data-readiness/): "Is our data ready for AI?" Collect, clean and label it so models can learn from it.
- [Sustainability reporting](https://dscvryai.com/use-cases/sustainability-reporting/): "What are our emissions, by site?" From raw figures to board-ready reports, every number auditable.

## Solutions (real work)

Outcomes as reported at the end of each engagement. Clients are not named, except Selati Game Reserve, named with its permission.

- [Production loss](https://dscvryai.com/solutions/manufacturing/): Automotive (India · global): 3 days → ~7 seconds to produce the production loss analysis; 31,190 claims reconciled, each with a justification pack; ~1,728 hours a year reclaimed.
- [Document operations](https://dscvryai.com/solutions/logistics/): Logistics (India · global 3PL operator): $2M+ revenue recovered; 80% fewer accounts-receivable exceptions; ~14 days → minutes processing time; 40,000+ documents a month.
- [Financial control](https://dscvryai.com/solutions/finance/): Apparel (South Asia · global): order-book questions answered in under 60 seconds, previously ~3 days. Building products (North America): proof of concept progressed to enterprise rollout.
- [Legal intelligence](https://dscvryai.com/solutions/legal/): Food services (multiple countries): 82% shorter review cycle.
- [Compliance](https://dscvryai.com/solutions/compliance/): Real estate (India): ~2 years → ~3 months delivery timeline.
- [Industrial data](https://dscvryai.com/solutions/industrial/): Chemicals (global): safety data sheet extraction API deployed in 3 days. Renewable energy (India): plant-level visibility in 15 minutes, every figure traceable to meter-level data.
- [Engineering evidence](https://dscvryai.com/solutions/software-engineering/): AI-assisted software delivery.
- [Sovereign operations](https://dscvryai.com/solutions/sovereign/): Restricted environments.
- [Wildlife conservation](https://dscvryai.com/solutions/conservation/): Selati Game Reserve, South Africa: concept proven, feasibility assessed.

## How we work

- Forward deployed engineering: senior engineers on the ground with your teams, from definition through to production. Delivered under our ISO/IEC 27001:2022 certified information security management system, with forward deployed engineering in scope.
- Strategy workshop ("What should AI do for us?"): the question every AI strategy should start with, answered and acted on.
- Engineers where you need them: a focused team embedded alongside your developers, for speed, capacity or specialist skills.
- The result, not the system: we deliver and operate a defined outcome against agreed service levels.
- Details: https://dscvryai.com/how-we-work/

## Deployment

Ours, yours or third-party: we deploy any model or product, wherever the work requires. Hosted (managed by Dscvry), public cloud (AWS, Azure, Google Cloud, Oracle), private cloud or VPC, on-premise, edge, offline, and air-gapped (physically isolated). The same controls apply in every mode.

Details: https://dscvryai.com/deployment/

## Insights

### Articles
- [The best AI strategy is sometimes no AI](insights/best-ai-strategy-is-sometimes-no-ai/): A client wanted AI in a workflow with fixed rules. We told them not to: ordinary code was cheaper, faster and easier to audit.

### News
- [ISO/IEC 27001:2022 certified](insights/iso-27001/): Our information security management is certified, with forward deployed engineering in scope.
- [SOC 2 Type II](insights/soc-2-type-ii/): Independently audited for security, availability and confidentiality.
- [DotPlot is now Clarify](insights/clarify/): Ask your enterprise data questions in everyday language, within each person’s access.
- [EcoWatch at Selati Game Reserve](insights/ecowatch-selati/): Automatic species identification from camera-trap images, built to run offline: a proof of concept at Selati Game Reserve.

### Research
- [A small model that reads NDAs](insights/nda-clause-model/): It rules on 17 standard clauses and quotes its evidence: 84.5% on a sealed test set, running on an ordinary CPU.
- [OCR only what needs it](insights/what-needs-ocr/): Find the parts of a page that exist only as pixels. We read 5.5% of the page area instead of 98.8%, with fewer word errors.
- [Every word, straight from the PDF](insights/where-is-the-text/): No model needed: 99.55% of 1.7 million held-out reference words found, each with its box.
- [Scan or text, in a millisecond](insights/scan-or-text/): A 116 KB classifier that tells whether a PDF needs OCR without rendering it. It got all 525 test files right.
- [What is this file, really?](insights/what-is-this-file/): A 39 KB check that tells what an upload really is, and whether it is locked, renamed or cut off. 1,914 of 1,917 held-out files right.
- [Everything a page draws, boxed](insights/where-are-the-regions/): Every text line, image, shape and annotation on a PDF page, read straight from the file. On 967 test pages the boxes hold every pixel of ink.
- [When a spreadsheet changes its meaning](insights/blind-spreadsheet-benchmark/): Twenty real public spreadsheets, three editions each. In 13 of 40 edition pairs, an automated reader should stop.

### Evaluations
- [TypeSafe’s Jev, tested on real work](insights/jev-eval/): 97.7% right on 177 scored questions, including invoices, expenses and a plant restart with an expired certificate.
- [One sentence fixed Jev’s one flaw](insights/jev-eval/): Asked two related questions with no rule, it contradicted itself in six of eight cases. Stating the rule fixed it.
- [Cloudflare’s Clef, close behind](insights/clef-eval/): 94.9% on the same 221 questions, and it caught 27 of 30 requests that needed a person to sign off.
- [Clef-Flash misses the sign-offs](insights/clef-eval/): The faster model scored 86.4%, and caught only 17 of 30 requests that needed a person to approve.

## Questions

### What does Dscvry do?
We build and deploy models, agents and intelligent systems that automate complex work for industry, government and defence. Each system is configured around your data, rules and operations, and tested on your own work before you rely on it.

### What do you build?
Models (our own, yours, or third-party), agents that do multi-step work and stop for a person, ready-made products such as Structurify, Clarify and GRID, APIs that plug it all into your systems, and the data pipelines behind them.

### Can you deploy third-party models and products?
Yes. We test new models on real work, publish the results in our Insights, and deploy the ones that fit alongside our own models and yours. We also deploy tools you have already chosen, inside your boundary.

### Where can it run?
Wherever the work requires: hosted by us, in your public cloud (AWS, Azure, Google Cloud or Oracle), in a private cloud or VPC, on your own servers, at the edge, offline, or fully air-gapped.

### How do you work with clients?
You choose how involved we are: a strategy workshop on what AI should do for you, engineers embedded with your team, forward deployed engineering from definition to production, or a defined outcome we deliver and operate against agreed service levels.

### Which industries do you work in?
Manufacturing, logistics and supply chain, finance, legal, compliance and ESG, software engineering, industrial and energy operations, conservation, and government and defence: work where decisions carry consequences.

### Are you independently audited?
Yes. We hold a SOC 2 Type II report covering security, availability and confidentiality, are certified to ISO/IEC 27001:2022 with forward deployed engineering in scope, have been independently assessed against the EU GDPR, and are penetration-tested every year by an independent third party.

### Where are you based?
London, through Discovery AI Limited, and Hyderabad and Chennai in India, through ReDscvry Technology Private Limited, with engineering and delivery across both countries.

### How do we start?
Bring us one problem where being wrong is expensive. We work out the data, models and agents it needs, test it on your own work, and agree where it should run.

## Security and compliance

- SOC 2 Type II report from an independent auditor (security, availability and confidentiality).
- ISO/IEC 27001:2022 certified information security management system, with forward deployed engineering in scope.
- Independently assessed against the EU GDPR.
- Penetration-tested annually by an independent third party.
- Details: https://dscvryai.com/security/

## Company

- Discovery AI Limited, 167-169 Great Portland Street, London W1W 5PF, United Kingdom.
- ReDscvry Technology Private Limited, Hyderabad, Telangana, India. Team also in Chennai.
- Podcast: Let's Talk AI (https://open.spotify.com/show/0ITRBoHWxC0oVxbM35g4Wc)
- Careers: https://dscvryai.com/careers/
- Contact: https://dscvryai.com/contact/ (hello@dscvryai.com)