Accurate, complete, and effectively unreadable: the standard fate of internal
documentation. Questions went to whichever colleague was assumed to know.
ClientGreek retail group
Users5–10 finance staff
StatusLive · extended to a second company
DeliveredSame engagement as the HR system
The problem
Four to five hundred pages of internal accounting manuals. The content was correct and
thorough. It was also, in practice, never opened, because finding the one paragraph
that answers your question is slower than asking the colleague two desks away.
So people asked the colleague. And the answers varied, depending on who was asked and
what they remembered.
What I built
A retrieval assistant over the manual corpus, returning natural-language answers grounded
in the actual policy text, and always showing which document the answer came from.
For policy content, citation isn't a feature. It's the product.
An answer a finance employee cannot trace back to the source isn't merely unhelpful.
It's a compliance risk. The assistant links to the document it drew from, so the person
asking can confirm it before acting on it.
Architecture
Delivered in the same engagement as the candidate-matching system, for the same client,
and deployed into the same secured environment: separate resources per application,
one hardened network boundary. Both the cost argument and the security argument point
the same way: one well-defended environment is cheaper to run and smaller to protect
than two.
Corpus · managed by the client
1
Manuals400–500 pages, added and updated by the client
indexed
2
Knowledge Basepolicy text, kept separate per company
Shared secure environment
3
Same private subnetseparate resources, one hardened boundary
question
4
Bedrock agentanswers grounded in the policy text
always with a source
5
Cited answera link to the document it came from
The client adds and updates their own manuals. When a policy changes, they change it.
Each company's policy text is indexed separately, which is what let the same system serve a second company without touching the first.
It runs in the same secured environment as the candidate-matching system, with its own resources behind one hardened boundary.
Answers are grounded in the actual policy text rather than generated around it.
Every answer carries a link back to the source document, so a finance employee can confirm it before acting on it.
Shared secure environment, isolated resources, a separate index per company.
The outcome that mattered
The system is still in daily use. More tellingly, the client came back and asked for it
to be extended so that another company in their group could use it for their own
manuals.
Delivery metrics are what a consultant claims. A client asking for more is what a client
does.
Nobody requests an extension to a tool they've quietly stopped opening. It also proved
the thing was more general than the brief: built for one company's accounting manuals,
it transferred to a different business with an entirely different set of documents.
Every organisation has manuals nobody reads. This one turned out not to be a project but
a product.
Open to remote roles worldwide
Athens-based Solutions Architect building and running production GenAI on AWS. Remote worldwide, or hybrid in Athens.