Alexandria

What it does

A document scanner drops PDFs onto a network share. From there the pipeline is automatic:

1. Read it. Scanned pages are images, not text, so every page goes through on-device OCR.
2. Classify it. A local language model decides what kind of document it is — a bill, a notice, a statement, junk.
3. File it. The document is renamed and filed where it belongs.
4. Extract the obligation. If there is a due date or an amount, it becomes a structured fact with a date attached.
5. Surface it. A morning briefing lists what is due, what changed, and what needs an answer.

Alongside the mail pipeline it keeps a searchable index of personal documents, triages email into what actually needs a reply, and proposes calendar entries and tasks.

The decision worth explaining

The first version asked the language model to decide whether something deserved an alert. It was wrong in a way that took a while to see: the model would read a notice with no due date printed on it and produce one anyway — a plausible date, correctly formatted, entirely invented. It would also give different answers to identical inputs, even with the sampling temperature set to zero.

So the alerting logic moved out of the prompt and into a deterministic parser with tests around it. The model still does what it is genuinely good at: turning a wall of scanned text into a short, readable summary. The part that wakes me up is code.

That split is the whole design philosophy, and it is the part I would want to talk about in an interview. A system that is confidently wrong about a deadline is worse than no system, because you stop checking.

Privacy was a constraint, not a feature

This thing reads tax documents, medical bills, and bank statements. That ruled out every hosted model before the first line was written.

• Inference is local. Models run on hardware I own. No document text is sent to an external service.
• OCR is on-device using the operating system's own vision framework — fast, accurate, and offline.
• Nothing writes to my calendar without me. The assistant proposes; I approve; then it acts. There is no path where it books something on its own.
• The API surface is tested for it. There is a test module whose only job is to check that the endpoints fail closed.

Deadline tracking

The module I use most watches for job-posting notification emails and turns them into tracked items with their real filing deadlines, a submission checklist, and a flag for whether I am actually eligible to apply. It is a small, unglamorous parser. It is also the reason I have not missed a filing date.

The same machinery handles bills: a due date extracted from a scanned notice becomes a dated fact, and dated facts show up in the briefing before they are late rather than after.

How it was built

I specified and directed this system, working with AI coding agents to implement it — I own the architecture, the guardrails, and the decision about what counts as done. The front-end work across my projects is my own; for Python services like this one, my contribution is the design, the constraints, and the judgment about what the system is allowed to do. That distinction matters to me, so I state it rather than letting a reader assume.

The discipline that makes the approach work is the same one that makes any delegation work: know what you are asking for, know what correct looks like, and do not accept output you cannot test.

Honest limits

• It is built for exactly one user on one machine. It is not multi-tenant and was never meant to be.
• The local models are small. They summarize well; they are not asked to reason about anything consequential, for the reason described above.
• OCR on a bad scan is still a bad scan. Low-contrast and skewed pages are the usual failure.
• Classification of unusual documents needs correcting more often than I would like. The filing is a suggestion I can override, not a verdict.

Seeing the code

The repository is private — it is wired into my own accounts and documents, so publishing it is not something I can meaningfully do. If you are evaluating me and want to look at the code or walk through the architecture, ask me and I will arrange read access or a screen share.