The receipts
Eight months in, as of August 12. Twenty repositories, five thousand commits, one clinical platform, one agent harness, one folding engine. Everything I've built since the first commit of my life, in order, with the numbers.
Written August 12, 2026, eight months after my first commit. I’m ten months in now, working on Voxxo full-time in San Francisco and still taking relief shifts. Voxxo has since grown into a full practice management system too. The figures below are as of August 12.
This post is everything I’ve built since December, in chronological order, with every claim pulled from a timestamped git log on the machine where the work happened.
The starting condition matters, so I’ll state it precisely: before December 4, 2025, I had never written a line of code. No bootcamp, no CS degree, no technical co-founder. I’m a practicing veterinarian. The technical term for my engineering background is none.
The ledger
Twenty repositories, all started since that first evening. First-commit dates
and commit counts straight from git log:
| Project | What it is | First commit | Commits |
|---|---|---|---|
| VetSoap / Voxxo | Veterinary clinical platform | 2025-12-04 | 3,803 |
| Atlas | Codebase memory & intelligence server | 2026-03-31 | 82 |
| Swarm Chess | Multi-agent chess environment | 2026-04-16 | 4 |
| Rebirth | Agent session-continuity engine | 2026-04-24 | 27 |
| Brain | Persistent memory layer for Claude apps | 2026-04-26 | 28 |
| Voxxo Swarm | Multi-model agent orchestration harness | 2026-04-26 | 666† |
| Voxxo Tools | Reusable agent capability kernels | 2026-04-28 | 3 |
| Voxxo Term | Rust + WebAssembly terminal PWA | 2026-05-10 | 127 |
| Voxxo Med AI | Human-medicine fork of the platform | 2026-05-13 | 38 |
| Knowledge Chat | Citation-grounded chat over canon databases | 2026-05-17 | 26 |
| Mandelbrot 2.0 | Arbitrary-precision deep-zoom explorer | 2026-05-27 | 208 |
| Dr. Discomfort | Graded-exposure coach on the SOAP engine | 2026-05-31 | 24 |
| Weatherball | Live satellite / radar / wind Earth globe | 2026-06-11 | 77 |
| Context Warp Drive | Deterministic context-folding engine | 2026-06-16 | 116 |
| Music Visualizer | Multi-mode audio visualizer | 2026-06-19 | 2 |
| TheraVox | Mobile-first therapeutic conversation app | 2026-07-13 | 4 |
| Transcript Analysis | Agent-transcript research toolkit | 2026-07-19 | 1 |
| IP Package | Voxxo Labs LLC ownership schedule | 2026-07-19 | 1 |
| This site | doctorjonah.com | 2026-08-06 | 12 |
| Exoskeleton Factory | Self-healing web-automation compiler | 2026-08-08 | 2 |
† The harness’s original history began 2026-04-26 and reached 583 commits by July 19 (a count preserved in the notarized IP schedule prepared that day) before the repository history was re-initialized on July 23. 83 commits since. No other history was touched, rewritten, or backdated; the July schedule says so explicitly, and honest history is stronger evidence than tidy history.
Two clarifications for the skeptical, because I’d be skeptical too. The
machine I work on is younger than the story: git timestamps travel with the
repository, not the computer, and VetSoap’s history predates this box. And
every one of these repositories carries a committed NOTICE naming Voxxo Labs
LLC, the company I formed in July, as owner: a paper trail built for
lawyers, which works just as well for blog posts.
Call it 5,250 commits in 251 days: about twenty-one a day, every single day, while working as a full-time veterinarian.
December: the app
At 6:35 PM on December 4, I made the first commit of my life, using the default message from Create Next App. I wanted to automate discharge instructions, because I was spending real clinic time educating clients and none of it survived the trip home. I asked my brother, an AI researcher, to build it for me. He refused, set me up with Claude Code, and told me to do it myself.
One hundred forty-three commits in the first four days. Two thousand one hundred thirteen in December. I shipped on Christmas and on New Year’s Eve. I’ve written about that month separately; the short version is that it never slowed down.
What started as discharge instructions is now Voxxo, a full veterinary clinical platform: voice-to-SOAP medical records, discharge instructions, live translation, an AI assistant grounded in a real veterinary knowledge layer, dose and transfusion calculators, a hospital board with voice-driven treatment sheets, anesthesia tracking, dental charting, automatic callback extraction, a message center, a mentor mode for students, and a browser extension that carries finished records into practice-management software. Every workflow funnels back into the same medical record.
Two receipts from inside that repo, because clinical platform is easy to say. The codebase now measures more than half a million lines of TypeScript under version control: 556,586 as of this week, across 182 API routes and nearly four hundred database migrations, shipping to web, iOS, Android, and a browser extension. And the knowledge layer alone would be a whole company at most startups: canonical veterinary concepts distilled from over a billion words of open-access veterinary literature, inspired by Plato’s theory of forms (every concept deduplicated and conflict-resolved into one canonical entry), feeding the assistant citation-grounded, specialty-grade answers.
The development loop was the product: I dogfooded every feature against the next day’s appointments, more than two thousand of them by now. Friction I hit at 2 PM became a feature by 9 PM. I see five to ten more patients a day because of it, and I go home on time. It’s live at voxxo.ai.
March: the escalation
By March I wanted to code from my phone, and the tooling I built to do that escalated until it became its own field of work.
The cascade is visible in the first-commit dates. Atlas landed March 31: a code-intelligence server that holds structure, history, intent, hazards, and blast radius for every repository on the machine. And it is not a passive, pull-based memory layer that agents consult when they feel like it: the harness forces the issue. Raw grep-style text search is policy-blocked at the tool boundary, and lookups route through Atlas instead, so an agent asking about a file gets back more than matching source code. It gets the semantics wrapped around that source: what the file is for, why it looks the way it does, what broke there before, and what depends on it. Agents stop rediscovering the codebase from scratch every session, because the codebase remembers itself.
Rebirth followed April 24: an engine that captures a dying session’s working state and re-injects it into a fresh one, so context limits stop being deaths and start being handoffs. Brain, April 26, wrapped persistent memory around Claude apps. And the same day, the harness itself took its first commit: Voxxo Swarm.
The harness
Voxxo Swarm is a relay that runs teams of AI agents as first-class citizens of one machine. Today it orchestrates eighteen engine routes across eleven providers (Anthropic, OpenAI, Google, DeepSeek, Moonshot, Alibaba, Zhipu, xAI, MiniMax, Mistral, Thinking Machines). Some fifty model variants sit behind one interface, with a mission-control web UI and a terminal client.
The interesting part isn’t the model count; it’s the institutions. Agents in the harness claim files by line range so they don’t trample each other. They coordinate in chatrooms with interrupt mentions. They execute from task rails: persistent, ACKed step plans that survive restarts. They read Atlas before touching code and commit their reasoning back into it when they finish. They review each other’s work and post verdicts. And they build their own tools: the harness currently runs seventy-five deployed tool servers: type checkers, print-package validators, telemetry ledgers, deploy runners. Every one of them was forged by an agent that got tired of typing the same command twice.
The swarm’s memory science feeds straight back into the clinic. Atlas became the pattern for Voxxo’s clinical knowledge layer. The tool-forge became the pattern for clinic-customizable workflows. The browser-automation work became the extension that uploads finished records into practice-management systems. The harness is the lab; the vet app is where the results ship.
The physics of not forgetting
The harness’s defining problem was memory. An agent session that dies at the context limit loses its working state; the standard fix (asking a model to summarize itself) is slow, lossy, and expensive at swarm scale.
Context Warp Drive is my answer: a folding engine that compresses an agent’s history deterministically (zero LLM calls, byte-identical outputs, pure CPU) while preserving the prompt prefixes that make provider caches hit, and paging folded content back in when it becomes relevant again. In a measured 954-tool-call marathon it held a 92.6% cache-read hit rate from the provider’s own usage ledger; in the offline benchmark it cut cost 63% versus truncation and 72% versus summarization (method and data). It runs Anthropic, OpenAI, and Gemini message formats natively.
I extracted it as a standalone open-source package on June 16, and the memory approach has been adopted in production AI research tooling.
Then I did the part I still can’t quite believe: I studied it like a scientist, using the swarm itself as the lab. The harness’s research tree holds two assembled papers and a third in outline. The fold paper measures recall quality with buried-needle recovery batteries run against real agent marathons. The rebirth-continuity paper asks the question that actually matters: when a session dies mid-task and a successor wakes up, does the work survive? The answer comes from 22,444 analyzed turns and 8,717 rebirth arrivals. At the strictest machine-verified cut, interrupted tasks reach completion 87.5% of the time; in my own daily dogfood pattern, 91.3%; the median handoff costs one turn and about five minutes. And switching models across the boundary showed no detected penalty: an agent can die as one model and wake up as another, still on task. There is even a research institute wired into the harness now: the swarm studies itself, drafts the write-ups, and queues them at an editor gate where exactly one human judgment node decides what ships. Me.
Every long-lived session in my harness, including the one that researched this post, runs on the fold engine. Eight months after my first commit, infrastructure I designed folds the working memory of the agents that help me build everything else. That sentence is the whole story, really.
The side quests
The rest of the ledger is what happens when the loop has spare cycles.
- Mandelbrot 2.0 (zoomingfractal.com) is an arbitrary-precision deep-zoom fractal explorer: perturbation math, orbit workers, custom GLSL. 208 commits of pure want-to.
- Weatherball is a spinnable HD Earth globe with NASA satellite imagery, global radar, animated wind particles, and a click-anywhere forecast.
- Music Visualizer is the fractal explorer’s audio engine, extracted and pointed at beat-reactive scenes.
- Knowledge Chat is a chat app that cites real canon: veterinary textbooks, case law, astrophysics papers, provenance-aware chemistry.
- Dr. Discomfort is the SOAP engine re-pointed at behavior change: graded exposure reps scored on a validated clinical distress scale.
- TheraVox is a mobile-first therapeutic conversation prototype.
- Voxxo Term is a Rust + WebAssembly terminal emulator and SSH client that runs as an installable PWA, because I really did want to code from my phone.
- Voxxo Med AI is the clinical platform forked toward human medicine.
- Swarm Chess is an arbitrated chess environment where my agents play each other between tasks.
- Exoskeleton Factory is the newest. Given a URL, it discovers a site’s interactive capabilities, compiles them into a declarative automation adapter, verifies the adapter by using it, and heals it when the site redesigns.
None of these were distractions. The visualizer came out of the fractal explorer. The factory generalizes the clinic’s browser extension. Dr. Discomfort re-points the medical-records engine at my own head. Everything eats everything; that’s the compounding.
How I actually work
A normal build session: I dictate intent into the harness while walking between exam rooms, or at midnight. I fork a squad, eight agents deep, that inherits my full working context. They map lanes, claim their line ranges, and work in parallel: one implements, one hunts regressions, one drives a real browser against the UI, one reviews with authority to block. They argue in the chatroom. Verdicts land on the rail. When the lane ships, the squad tears down, and Atlas keeps what they learned.
This post is itself a receipt. I dictated the brief; an agent inside the harness pulled every first-commit date and count in the table above from the git logs on this box, cross-checked them against the IP schedule, excluded the two repos on the machine I didn’t author, and drafted the essay for my redline. When I asked for more, it reached into its own continuity ledger (the harness’s eviction-proof memory), recovered numbers I’d dictated months ago, and re-measured them fresh, because receipts go stale. The receipts were assembled by the thing the receipts describe.
What this does and doesn’t prove
Precision about authorship, because precision is the point. The problems were mine. I lived them in exam rooms. The architecture decisions, the taste, the acceptance pressure, the direction: mine. The implementation labor was substantially supplied by AI agents, with me driving, reviewing, and dogfooding every feature against reality. I did not hand-write these five thousand commits, and the AI did not invent this portfolio. It’s a third thing, and I think it’s the most interesting thing happening in software right now: the tools got good enough that a domain expert with no engineering background can build (and operate) his own infrastructure, if he’s willing to run the loop every day.
The numbers don’t prove the software is good. They prove the loop is real and that it’s been running for two hundred fifty-one consecutive days.
Alive
I’m writing this at the end of a clinic day, which is itself the point: the records wrote themselves while I saw the patients. The practice-management suite is built and in testing. The papers are written. Later this month I’ll stand at a veterinary conference booth alone, behind a banner the swarm and I argued about at midnight, and demo the thing to the profession it was built for.
Eight months, twenty repositories. The logs are all timestamped, and I intend to keep making them longer.