How this works

AI assistant for Dhavan’s published portfolio material. Answers can be incomplete or wrong, so do not submit confidential code or data.

Your questions go through Hugging Face, and a model run by DeepInfra writes the answers. Chats may be retained for about 30 days to operate, debug and improve the assistant; DynamoDB TTL deletion is approximate. Cal.com handles call bookings. Details are in the privacy policy.

Use the suggested questions or ask about published projects, credentials, and engagements.

Projects / Agent Brain

Answers that cite their source, or say they don’t know.

Agent Brain is internal tooling I adapt in engagements, not a product you buy. It only answers from sources the user may read, and says “I don’t know” when the evidence is thin.

Dhavan Shah You work with the person who built it. The first 30 minute call is free, and if I can’t help, I say so before you pay anything.

Governance checksEvery change
Sources
Default deny
Answers
Citations required
Citation coverage too low
“I don’t know”
Conflicting facts
Flagged at ingest

Cross source leaks0

In a fuzz suite that checks an answer never draws on a source the user may not read.

Agents invent company facts when memory is weak.

A company assistant can make things up or leak context when retrieval, reasoning and permissions are treated as one loose step.

  • “Can we trust what it says about our company?”

    Answers tied to evidence

    Claims point back to the supporting material they came from.

  • “What happens when it doesn’t know?”

    Cite or abstain

    An answer is returned only with citations to retrieved sources. When coverage is too low, it’s replaced with “I don’t know.”

  • “Could it leak another team’s data?”

    Access aware by design

    What a person can see shapes what can enter an answer, and a fuzz suite checks it on every change.

  • “What if two documents disagree?”

    Contradictions flagged

    Conflicting facts are flagged at ingest for review, not merged silently.

“Not enough evidence” is a designed outcome, not a failure.

How it works: separate steps you can see.

Retrieval, reasoning, citation, abstention and access filtering are separate control points, not one chat over a document dump.

1 · Retrieve

Only what they may read

Sources are default deny, so what a person can see shapes what can be retrieved.

2 · Check

Evidence before prose

Writing the answer is a controlled step, and teams can see what was used and what was held back.

3 · Answer

Or say it can’t

It returns an answer with citations, or says there isn’t enough verified information.

What runs on every change.

Governance checks, not a one time review.

  • Permissions: sources are default deny, and a fuzz suite checks that an answer never draws on a source the user may not read, with 0 cross source leaks.
  • Cite or abstain: an answer is returned only with citations to retrieved sources, and is replaced with “I don’t know” when citation coverage is too low.
  • Contradictions: conflicting facts are flagged at ingest for review instead of being merged silently.

A known limitation

Benchmark scores are not published here. They will be run again with an untuned protocol before any figure is shown.

Where it fits

Company Q&A, agent context packs, meeting prep and policy aware assistants, especially when invented answers or data leaks across teams are unacceptable.

Before you get in touch.

Can I buy Agent Brain?
No. It’s internal tooling I adapt in engagements. The contract it follows, cite, abstain and respect access, is what a Production Pilot puts in place.
Does every engagement use it?
No. It’s for work where retrieval and permissions are the problem.
Who built it?
I did, around a clear contract: cite, abstain and respect access. It isn’t an unbounded chatbot over a document dump.
How do we start?
Send a short description of the workflow and what is blocking it. I reply within one business day. The first 30 minute call is free, and there’s no obligation until you approve the Diagnostic.

Tell me which company knowledge your AI must not get wrong or leak.

The first 30 minute call is free, and I reply within one business day.

Book a free call

How engagements work · Back to Projects