How I work with AI — the operating system, not the demo
Every case on this site applies one thesis to a product. This one applies it to me: how my own career, thesis and job search run on deterministic sources of truth, adversarial review, eval gates and automation with stop rules.
The thesis behind every case on this site — deterministic where it matters, models at the edges, evals as the definition of done — is not something I apply only to payment routing or nutrition parsing. It is how I run my own operation: career planning, a master’s thesis, a job search, this site. AI is in the loop everywhere. It decides almost nothing. This page is the receipts.
Facts live in one deterministic place
My CV is not a document; it is a build artifact. Every fact lives in a single versioned JSON file — English and Spanish in mirror — and a build script generates the four Word/PDF variants from it. A guard script refuses to sync the website if any artifact is older than its source, and names the exact command that was skipped. The model that helps me edit the CV never edits a PDF; it edits the source of truth, and the pipeline does the rest. Same shape as the routing engine: the hot path is deterministic, and no amount of model eloquence can push a stale fact to production.
The career plan splits the same way. The why — thesis, decisions, assumptions, risk register — lives in versioned markdown. The where are we lives in a database the AI copilot can query. And the write rules are explicit: the copilot never writes an interpretation without showing me the exact change first, but it records consummated facts — an application sent, a rejection received — autonomously, reporting afterwards what it wrote. Recording facts is mechanical; deciding what they mean is not. That line, written down, is what makes delegation safe.
An AI never grades its own work
My master’s thesis deliverables go through an adversarial cycle before anything is submitted: one model authors, then separate reviewer roles — a professor simulation scoring against the official rubric, a methodological examiner, a citation auditor — attack the draft, a corrector applies the surviving findings, and a different model grades the result. The author never grades itself; findings must survive verification before they cost me an edit.
The cycle earns its tokens. The examiner caught a quotation presented as literal that was actually my own translation — an APA violation a human reviewer would flag in a defense. The auditor verified 26 of 29 figures against primary sources, word by word where it mattered. Recent deliverables came back from the real evaluation at 10/10 and 20/20; the write-ups scored 15.5/16 before submission, with the last half-point being a missing word the final review pass caught. Plausible-but-wrong is the failure mode of working with generative models; independent adversarial review is the only defense I trust, because it is the one that regularly finds things I was sure were fine.
Evals gate every model change
Nothing with an LLM in it ships on vibes. The nutrition app’s extraction chain is gated by a golden-set eval harness that turned “the prompt feels better” into a scored regression gate — and demoted one model’s baseline to a smoke test when the eval proved its failures were noise, not signal. The payment orchestrator’s single LLM edge is gated by an eval that counts hallucinations instead of wishing them away, with a schema, an enum gate and a safe fallback underneath. The rule is the same everywhere: a model change without an eval is a mood, not an improvement.
Automation needs a stop rule
My job search runs on autopilot three times a day: an agent scans new postings, scores them against a written two-track rubric — domain fit and career trajectory, with explicit gates for compensation band and async-first policy — and stays silent. It interrupts me only when a posting clears the “perfect job” threshold; everything else waits for a single morning digest. The design decision is the silence, not the scanning. An automation that pings you constantly trains you to ignore it — the same lesson as the eval whose false alarms you learn to ignore. Alerts are a budget, and almost everything fails the bar.
Tokens are unit economics
Working agentically all day makes token spend a real cost line, so I treat it like one. The driver is not the size of a change but the length of the mechanical queue behind it — builds, deploys, verification logs — so judgment work happens in the expensive model’s context and mechanical queues get delegated to cheaper models, whose logs die in their own context instead of accumulating in mine. Compressed interaction modes strip the pleasantries from every exchange. I learned the price of not doing this by measuring it: an early fully-orchestrated academic deliverable burned half a million tokens on the semester’s easiest activity. The pipeline that came out of that autopsy produces better work at a fraction of the cost — because the roles are narrower and each one reads only what it needs.
What stays human
Every consequential choice in the system is a numbered, written decision with a date, a rationale and the assumptions it rests on — thirteen of them so far steering the plan, each one reviewable when reality disagrees. Models draft, verify, record and remind. They do not decide what the niche is, what a fact means, or what gets published under my name; every piece of public content clears an explicit approval, the same way nothing here wrote itself into production. That is the actual skill I am selling — not using AI, but knowing where it earns its place. The products are the proof; this page is the practice.