Controlled Copy: answers you can trace back to the source
Company: Personal project, built as a work sample for a job application
Duration: October 2026, four evenings in one work week
Tools & Technologies: Python, FastAPI, htmx, SQLite, Docker, OpenRouter, hybrid retrieval (keyword and vector search)
Links: GitHub · Live demo offline for now, available on request
Embedded Files
Project Overview
Project Overview
Context
A small, solo project. I built it on four evenings during a regular work week as a work sample for a job application. It's a NotebookLM-style notebook: add documents, ask questions, and every statement in the answer links to the exact passage it came from. I added one thing NotebookLM doesn't have: document control.
Key Result
The server checks every quoted passage against its source and drops any statement it can't verify. When the sources don't cover a question, the app says so instead of guessing.
Tools
Python · FastAPI · htmx · SQLite · Docker · OpenRouter
The Problem
The Problem
In operations and quality work, the wrong document is worse than no document. An outdated revision, a draft or another site's instruction can look perfectly reasonable in an AI answer.
A general-purpose chatbot has no idea which version is the approved one.
The Solution
The Solution
Every answer cites its passage. A follow-up is rewritten into a full question, and a question the sources don't cover gets a refusal instead of a guess.
Retrieval with an evidence floor. Passages are found by exact terms and by meaning. If nothing matches well enough, no model is asked at all.
Verified quotes. Every statement must quote a passage word for word, and the server removes any statement whose quote isn't there.
Document control as data. Revision, status, effective date, site and role decide which documents may back an answer before the model writes anything. Old revisions and drafts are listed as not applied, with the reason.
Resolution Card: every required action is backed by an approved, current instruction for this site, role and date.
How It Was Built
How It Was Built
I wrote the specification and made every product decision. AI coding agents wrote most of the code; a second agent reviewed each change, and I reviewed and merged every one myself.
Before handing it in, I tested the live app by hand and triaged 23 findings into fixes.
Impact
Impact
500+ automated tests, in CI
6 of 6 questions answered correctly on a public 48-page document
13 of 14 document-control cases correct; the one miss chose the more cautious status
What's next
What's next
A code review after the project gave me a clear list, now open on GitHub: use the official SDK for model calls instead of my own HTTP client and retry logic, replace the remaining hand-written parts with proven libraries where they fit, name the architecture patterns in the docs, and add browser tests for more user journeys.
Why this Project Matters
Why this Project Matters
It shows how I'd put AI into a regulated operation, and how far one person can get in a few evenings with a clear spec: rules first, model second, every claim checkable.
Page updated
Google Sites
Report abuse