All case studies
Case Study · Thermal Power Generation · Energy & Infrastructure

Document Intelligence Platform for a Thermal Power Generator

Power purchase and fuel agreements, O&M contracts, regulatory orders and handbooks made queryable in plain English, with every answer cited and every query audited.

Document IntelligenceRAGOCRGuardrailsRBACAzure
Headline outcome10 weeksto go-live across four acceptance milestones

What the client was up against

A large supercritical thermal power generator already held a substantial data lake, but the unstructured half of its estate was effectively unqueryable. Power purchase and fuel supply agreements, O&M and EPC contracts, regulatory orders, petitions, arbitration records, consents, handbooks and compliance filings carried the obligations the business ran against, yet could not be searched, compared or cross-referenced. A question like which contracts carry a liquidated-damages clause linked to availability, and what each specifies, could only be answered by someone who already knew where to look.

Design, build and take live a Document Intelligence platform inside the client's own cloud tenancy, under the group's IT cybersecurity norms, across four acceptance milestones in ten weeks, then transition it into a twelve-month maintenance arrangement. Scope was frozen to document intelligence only, with financial analytics deliberately deferred to a later phase.

How we built it

01
Full Ingestion Spine
  • Native PDF, scanned PDF through OCR with layout and table reconstruction, Word, Excel and HTML
  • Automatic classification into a 25-class taxonomy agreed with the client's reviewers
  • Amendments, addenda and superseded versions linked to their parent so the current position is answerable, not just what the original said
Document IntelligenceAgentic Workflow Automation
02
Cross-Document Referencing With Cited Answers
  • A single question resolved against several documents at once, such as a PPA clause read against a regulatory order or a handbook procedure against a statutory consent
  • Within-document referencing across several-hundred-page contracts and their annexures
  • Every answer carries document name, version and page or clause reference, and the platform says so when the corpus does not support an answer
Document Intelligence
03
Guardrails Built Into Retrieval
  • Grounded answering only, with scope confinement so off-domain questions are declined rather than answered from general knowledge
  • Role-aware responses applied inside retrieval rather than as a filter on results
  • No derived computation, and document content treated as data so instructions embedded in third-party contracts are never executed
  • Unreadable pages flagged rather than guessed, with a full audit trail of every query, answer, citation and refusal
  • Each guardrail is delivered and verified as an acceptance test
Auto-QA AgentDocument Intelligence
04
Tenancy-Resident Architecture & Managed Delivery
  • Versioned Docker images on the client's own cloud servers with inference through its existing AI service keys, so no GPU is procured and no corpus data leaves the tenancy
  • Directory-based identity with three roles and MFA, managed secrets, private endpoints and security-log integration
  • An independent penetration test before go-live at our cost, a named delivery lead, fortnightly steering and a live risk register
Agentic Workflow Automation

Measured impact

0,000
pages across roughly 50 long-form contracts, orders and handbooks in the initial corpus, phased in three priority tranches
0
document classes in the taxonomy driving routing, access control and retrieval quality, with administrators able to correct classification
0
guardrails delivered and verified as acceptance tests, from grounded-answering-only to a complete query and refusal audit trail
0
GPUs procured and zero corpus data leaving the client tenancy, with consumption billed under the client's existing cloud agreement
Ten-week delivery across four milestones with a 15-day grace period for final sign-off, followed by twelve months of maintenance. Feedback from nominated reviewers is applied by people to retrieval tuning and prompts; no model is retrained on client content.

Ready to move beyond the proof of concept?

We work with a small number of enterprise clients each year. If you have a clear AI mandate and a production problem to solve, let's talk.

Structured Discovery, workflow mapping, data audit, ROI modelling. No commitment beyond the first session.