Case study · Defence Intelligence Assistant

Grounded, traceable AI over your document estate, safe to accredit.

An assistant that lets planning staff interrogate manuals, procedures and doctrine in plain language, with every answer grounded in source and traceable to the page. Accuracy and traceability are built into the architecture; security is kept deterministic by design.

See the live demo →Download the deck (PDF)

The demo runs on notional doctrine for illustration, not real classified material. Access requires a work email.

The problem · current state

The documents are digital. The way staff use them is not.

What changed

Manuals, procedures & doctrine digitised

All content now lives as PDF and DOCX in one central database.

A single source of record

No more physical textbooks scattered across desks and shelves.

What didn't

Staff still read to retrieve

Finding one planning factor can mean opening several 300-page documents.

Search is keyword, not meaning

A query only works if you already know the document's wording.

“Leverage AI over the new database to increase the efficiency and output of central planning staff, while accuracy and traceability remain key.”Chief of Staff

What “good” looks like

Four criteria the solution is measured against

constraint

Accuracy

Answers grounded only in retrieved source, no fabrication, no guessing.

constraint

Traceability

Every claim links to the exact document, section and page it came from.

outcome

Security

Users only ever see content their clearance and need-to-know permit.

outcome

Efficiency

Minutes of reading collapse into a sourced answer in seconds.

The solution

Ask in plain language. Get a grounded, cited answer.

A retrieval-augmented assistant sits over the document estate. It retrieves the passages a user is cleared to see, reasons only over them, and returns an answer with inline citations that open the exact source. The interesting engineering is in what is not left to the model.

Architecture

How it works

Ingestion runs offline; retrieval and generation run per query — Documents are chunked, embedded and stored with their access tags once, offline. Every question then runs the same online path: plan, gate, retrieve, rerank, and generate a cited answer. AI does the language work; deterministic software does the security and plumbing.

Cast a wide net, then narrow hard — Vector search retrieves the top 50 candidates for recall; a cross-encoder reranker rescores them and keeps the best 8 for precision. The model answers strictly over those 8 passages, emitting inline citations. Follow-up questions are rewritten against conversation history first, so “how is that calculated?” resolves to a fully specified query.

Access is filtered before retrieval, and it isn’t AI — The user’s identity (clearances and compartments) is intersected with each chunk’s ACL tags to produce the allowed set, the only candidates retrieval or the model ever see. It’s deterministic and fully audited. You wouldn’t want an LLM deciding who sees classified documents; that decision must be provable and repeatable.

What’s genuinely AI, and what’s deliberately not — AI is used where it adds judgement and recall: embeddings, reranking, grounded generation, and conversation memory. Authentication, access-control filtering, ingestion and the interface are deliberately classical software, the controls that sit around the AI, where correctness and security must be guaranteed.

The impact

Reading time becomes planning time

~30 min

to locate one sourced fact by hand

<1 min

to the same answer, with citations

Hours / week

recovered per planner, redirected to planning judgement

Consistency

the same sourced answer regardless of who asks

Illustrative pilot hypotheses, measured against DefTech's own baseline.

Delivery · deployment options

The same pipeline runs where accreditation requires

Start in cloud for the pilot, migrate on-prem without re-architecting.

Pilot fit

Dedicated cloud / VPC

Fastest path to a working pilot
Single-tenant, network-isolated deployment
Elastic scaling, managed updates
Ideal for an unclassified corpus to prove value

Lower setup effort · quickest to value

Classified fit

Private / air-gapped

Models run inside your security boundary, no data egress
Full control of data residency and the update cycle
Aligns with accreditation for classified material
Runs on your hardware or a sovereign environment

Higher setup effort · maximum control

Delivery · a phased pilot

Prove value first, accredit second, scale third

Phase 0

Discovery

Baseline current effort. Assemble the golden Q&A set. Confirm corpus and access-control model.

~2 weeks

Phase 1

Pilot

One directorate, unclassified corpus, cloud. Measure against the baseline.

~6 weeks

Phase 2

Accredit & secure

Move to private deployment. Onboard classified material under full ACL.

accreditation-led

Phase 3

Scale

Roll out across directorates. Expand the corpus and monitor evaluation continuously.

ongoing

Each phase has a clear exit gate, DefTech decides to proceed at every step.

The full picture

The solution-architecture deck

The complete walkthrough, problem, architecture, security, deployment, evaluation methodology and risk register.

Open the deck full screen Download the PDF

Try it

See the Defence Intelligence Assistant answer a live question

Ask it in plain language and watch it retrieve, reason and cite. The demo runs on notional doctrine, and access is protected by a work-email check, you'll receive a one-time code, then land straight in the tool.

Launch the live demo →Read the thinking behind it