Jonah Tarashansky / selected work / 2026
I’m a veterinarian who learned to code to solve a problem I lived every day. I built the clinical platform, then a whole AI development swarm I can manage from my phone. The work now spans open-source AI tools, a mobile-friendly command-line app for controlling remote computers, and an interactive fractal explorer.
01 Veterinary AI platform
Live product
A voice-first AI platform for the whole veterinary hospital. Speak, and it drafts the record and fills in the dental chart and treatment sheet. It adds clinical insights from a veterinary knowledge base, catches the follow-ups and learns how each clinician likes their notes written.
Documentation used to keep me at the hospital long after closing. I built Voxxo while practicing full-time (during, between and after appointments). It started as a scribe for records and discharge instructions, then grew into a clinical workflow tool for the whole hospital: the record, the medicine behind it, the treatment sheet, the dental chart, the invoice and the follow-up call. Most recently it became a full practice management system too, with a Practice Hub for clinics that want everything in one place.
Voxxo fits into the system a clinic already has: its Chrome extension works with any cloud-based practice management software. Web, iOS and Android apps, in use at paying veterinary clinics.
The learning loop
Lessons stay with each veterinarian, plus a shared practice-wide layer
Voice-first
Voxxo writes the record from the visit conversation, and the same idea runs through the hospital board. Clinicians speak; AI turns what they say into structured entries (teeth, doses, times, vitals) for them to confirm.
108 pocket depth 6 buccal, 309 mobility 2 furcation 1, extract 404Dental findings land on the chart, tooth by tooth.
Maropitant 1 mg/kg SQ q24h, LRS 60 mL/hr, Vitals q4h for 2 days at 2pmA treatment plan becomes scheduled flowsheet entries, with doses worked out from the patient’s weight.
Change carprofen to 2 mg/kg, cancel gabapentin, add metronidazole 15 mg/kg IV q12hPlan changes are staged as treatment cards to apply or undo.
HR 120, RR 14, SpO2 98, ETCO2 35, iso at 1.5, temp 100.2Anesthesia vitals are parsed into their fields; drugs and fluids are dictated the same way.
Vaccines, treatment templates, anesthesia protocols and new patients on the whiteboard take dictation too, and the live translator works in speech on both sides.
The draft record appears the moment recording begins. Each pause uploads that part of the visit in the background, and a recording can move to another device mid-visit. When you finish, transcription and your learned preferences load in parallel.
On by default, every patient record gets an advisory review from the knowledge base: differentials, treatment rationale, monitoring, prognosis and drug interactions, each labeled by confidence.
Tens of thousands of reference chapters and hundreds of thousands of evidence-linked facts, each carrying a confidence label. Clinical insights and the clinical assistant draw on it.
Voxxo picks up the callbacks, rechecks and other commitments made during a visit. It proposes tasks with due dates and the exact words that prompted them, plus active cases for rechecks, ready for the clinician to approve.
The clinician and client each speak their own language. Voxxo translates both sides in real time and speaks the translations aloud, preserving veterinary terms and drug names. Currently in admin preview.
Clinics don’t have to switch. The Chrome extension works with any cloud-based practice management software: it reads the schedule and patient history from the page and fills finished records and discharge instructions back into the clinic’s own forms.
Practice management
For clinics that want everything in one place, Voxxo is now a full practice management system too: schedule, charts, clients, billing, stock and reports, with the Practice Assistant beside every tab. None of it is required. The rest of Voxxo works alongside the system a clinic already has.
Smart Route plans mobile visit days. Practice Forge is the clinic’s own tool registry: new tools can be drafted from the assistant chat, and an admin reviews each one before it is installed. Travel health certificates are in admin preview.
02 Agentic development
My development platform
The model is interchangeable. The accumulated work is the agent.
Each agent is a durable identity made from its transcript, task state, codebase knowledge and cognition, not a disposable model process. I can move a running agent to a different model mid-task while preserving its identity and working state, or fork inherited-context clones and whole squads so parallel workers begin already briefed.
Context Warp Drive carries decisions, evidence and open commitments through context limits and restarts. Durable task rails keep execution explicit. Atlas gives every agent source history, dependency maps and exact snapshots; Forge lets them turn repeated work into tested tools they can reuse.
Shared cognition cards, chatrooms, file claims, collision detection and independent review rings make the fleet coordinate rather than merely run in parallel. Live usage, cost and context telemetry keeps the system observable across providers.
It runs at fleet scale: nearly 3,000 agents have worked in Swarm since late August, with up to 199 in a single day and 99 running at once on one machine. It’s built to run lean. Each agent works inside its own memory limit, and heavy builds and tests are admitted only when memory allows. Those safeguards date from when Swarm ran on a 12-core, 24 GB server.
From the dedicated Mission Control mobile app on my phone, I can switch projects, inspect agent panes, speak or type instructions, fork workers, change models and intervene in live work. The same agents can navigate real browsers and, under supervision, operate a remote computer through a secure SSH bridge.
This is the system I use to build Voxxo. Its strongest ideas flow back into the clinic: Forge-built reusable workflows, long-horizon memory and learning loops that retain each veterinarian’s documentation preferences.
Explore the tools I’ve releasedThe agent stack
03 Long-horizon memory
Open source · MIT
Agents that work for weeks without forgetting, and without paying full price to reread their own history.
Every AI agent works inside a fixed context window. When it fills, the usual fixes break something. Truncation throws away the work in progress. Asking a model to summarize itself costs another model call, loses exact details and rewrites the history the provider had already cached, so the next turn pays full price again.
Context Warp Drive is the engine I built instead. It folds history deterministically: no model calls, only CPU, and the same history always produces the same bytes. That one property lets an agent’s memory be long-lived and cheap at the same time. Providers bill cached input at a fraction of the normal rate, but only when the start of the prompt is byte-for-byte identical, and a deterministic fold keeps it identical turn after turn.
Most agents haul the whole haystack around to keep the needles. Context Warp Drive keeps the needles (file paths, IDs, the newest steps and the agent’s own reasoning) and leaves the hay in the barn. When a piece of hay matters again, it pages back in.
Swarm’s long-running agents run on it. I extracted it into a standalone MIT library that works with Anthropic, OpenAI-style and Gemini message formats, includes fold packs for Claude Code, Codex CLI and Gemini CLI, and ships with more than 900 deterministic tests. Brain MCP brings its rebirth to the Claude Code CLI.
How it works
The cache numbers below weren’t earned by letting context pile up. Agents compacted it over and over, deterministically, so the cache kept hitting while each agent worked in a small context without losing stride.
Cache figures are the token counts providers reported back, added up from Swarm’s usage ledgers rather than estimated: 42,562,290,016 of 46,245,952,638 input tokens came from cache (92.03%). Of the rest, 3.06 billion weren’t in the cache and 0.62 billion were written to it for later turns. The date window is fixed, so rerunning the measurement gives the same counts. Rebirth figures come from the rebirth-continuity paper: 22,444 turns and 8,717 rebirths across eight model engines. In its controlled test across five model families, rebirth matched a full-context summary at choosing the right next action. Benchmark method and data.
Open tools & experiments
The systems behind my workflow also stand on their own. These public repositories show the engineering directly.
Open source · MIT
Give agents a map of the codebase.
Local code intelligence with structural search, dependency graphs, snapshots and a place to record the reasoning behind a change. A reusable MCP server, CLI and programmatic API.
View Atlas MCP on GitHubOpen source · MIT
Long-horizon memory that keeps the cache hot.
Deterministic context folding with byte-stable prompt prefixes, recall and rebirth. The engine behind Swarm’s long-running agents, released as a standalone library with fold packs for Claude Code, Codex CLI and Gemini CLI.
View Context Warp Drive on GitHubOpen source · MIT
Swarm’s memory, inside Claude Code.
Atlas code intelligence, Swarm’s transcript search (Tap) and rebirth, packaged as one 23-tool MCP server for the Claude Code CLI. Install it, run brain setup, and Claude Code can search its own past sessions, query and record codebase history, and hand unfinished work to a fresh session that wakes up already briefed. It all runs locally on SQLite and an on-device embedding model, so you get these capabilities without adopting the full Swarm.
View Brain MCP on GitHubPublic code
A benchmark arena where models play each other.
I pulled it out of my Swarm harness, where it pits models against each other: each color gets its own provider and model, round robins balance colors, and an Elo leaderboard ranks the field. The harness also tests whether a model plays better stateless, in one long conversation or across rebirth handoffs, and it guards against either side seeing the other’s reasoning. The public engine gives the model the board every turn, checks each move and returns illegal ones with the legal list for up to three retries.
View Swarm Chess on GitHub04 Developer tools
A mobile-first terminal: SolidJS on top, Rust underneath.
Voxxo Term lets me connect to remote computers and run commands from my phone or browser. The interface is a SolidJS app. Underneath it, a terminal engine written in Rust and compiled to WebAssembly parses the terminal stream and draws it to a canvas. The Android app adds a native SSH client, and the controls are designed for touch as well as a keyboard.
05 Interactive exploration
Follow your curiosity. Keep zooming.
An interactive Mandelbrot explorer for the browser and the phone. Pinch to zoom, then keep going far past the roughly 1015 magnification where ordinary 64-bit math gives out. The camera tracks its position in arbitrary-precision BigInt arithmetic whose precision grows with depth. Past 2100 the renderer switches to scaled perturbation theory: a high-precision reference orbit is computed off the main thread, the GPU renders every pixel as a small offset from it, and a guard catches the glitches that appear at extreme depth.
Opens directly in the browser. No account needed.
The thread through it all
I began in December 2025 with no coding experience. I built full-time alongside full-time veterinary work, then moved to San Francisco in August 2026 to pursue go-to-market full-time. I still take clinical relief shifts: real appointments keep the product grounded.
Read the longer story