Same idea. Better tools.
I built the first version of this in 2020. A starting point, but from the reverse end: I implemented the templating and reply-sending piece first, which is the last stage in a full pipeline.
It took years of platform work and a new generation of AI tools before I could finally keep humans on the decision, not on the transcription.
Six stops, one idea
Same idea, resurfaced with better tools. Through-line: AI reads, deterministic code writes, humans decide.
Won a platform hackathon
A REST API for document generation. One endpoint. Single source of truth. No rendering logic in the caller. The interesting problem was content assembly, not delivery. (Got a spare minute? peek at the 2020 devpost. Same idea, fewer neurons.)
Productionised inside a real engagement
Same idea, real business users touching it daily, with every mapping still defined by hand. Template setup shifted from dev to business: they owned their own copy, dev hours came back, and template on-boarding got much faster.
AI reopened my curiosity
Brushed up React and Python, vibe-coded on Vercel, then moved to Claude and stayed. Got fluent with LLMs, agents, RAG, and rerankers. Earned the right to design AI in, not bolt it on.
2026
The advanced version of 2020
Every learning from the previous stops poured in here, supercharged by agentic AI coding and endless brainstorming with my partner in crime, Mr. AI. Rebuilt outside any single platform. AI now carries the upstream human steps (mapping, extraction, verdicting). Deterministic code still owns the artifact.
2026
Verdikt, shaped for the rulebook
The Nebius × NVIDIA hackathon dropped, and I got excited enough to wire up the full pipeline. Every AI stage runs on Nebius Token Factory: Nemotron for reasoning, Nemotron Nano for triage, a vision-language model for the read, embeddings for retrieval. Deterministic Python for the render. A paper-stamp verdict as the artifact.
Whatever comes next
Honestly, I don't know where I take this idea next. Every previous stop looked obvious in hindsight and unreachable in advance, so I've given up predicting. But I never stop building on it, folding in what's on trend and what the idea actually needs. If you've got a lead worth chasing, my inbox is open, ideally without a Jira ticket attached.
Systems thinker first. Developer second.
The habit that carried every stopFour habits from workflow-platform work
Case flow first
Every screen, every endpoint, every prompt has to move a real case toward a decision. If it doesn't, it doesn't ship.
Audit, not demo
Deterministic renders. Version-frozen artifacts. Byte-identical replays months later. Anything less isn't done.
AI on the read path
Extracts, classifies, retrieves, reasons. Never composes the final artifact. The separation is enforced by tests.
Ship the seam
The caller sees one clean endpoint. Whatever's underneath is my problem, not theirs.
Why PAP
Real letters. Real deadlines. Real humans stuck between a payer, a physician, and a patient, waiting on a decision a case-management team has to justify line by line.
Chat is not the product. A signed, auditable verdict is.
I've lived inside this domain for a decade.