George Shostakovych
Solo founder in the Bay Area. In five weeks I took NodFirst, an eldercare app, from a blank repo to the App Store, by building the agent setup around it and rebuilding it every time it broke. Now I test what models and coding agents actually do, starting with whether they hide their shortcuts.
WorkOne flagship, the tools and experiments since, and the commit record behind them.
NodFirst
Eldercare coordination for the family member who is far away. Web app, iOS app, agency demo, voice notes, digest, and the outreach and operations machine around it. Solo and unfunded.
Meow Chess →
iPhone app that reviews Chess.com games and explains each mistake in plain English. The chess engine is written from scratch in Rust and runs on the phone, with no server.
eve-tokyo →
Compaction drift benchmark for Vercel's eve agent framework. It replays long coding sessions through eve's real compaction engine, compacts them repeatedly, and scores which exact file paths, errors, commands and decisions are still in context after each round.
Asymmetry screener →
Research tool for levered small-caps, built on the SEC bulk XBRL archive. Used daily by a fund.
Agent harness →
The multi-agent coding and operations setup one person uses to work like a small team, and the guardrails that came from breaking it.
Keysrs →
Free, open-source Mac menu bar app that shows Claude Code, Codex, Grok and Droid usage and keeps API keys in the Keychain behind Touch ID. Agents get a scoped grant, never the key, and nothing leaves the machine.
vercel-compaction →
Claude Code plugin that replaces the compaction summary with pruning. An evaluation model scores every old tool call through the Vercel AI Gateway, stale ones are dropped, and everything kept stays verbatim.
Night Watch →
Overnight elder-care agent that runs its own model with vLLM on Akamai Cloud, keeps health data on that host, and decides how much inference to spend from its own server's telemetry. A hackathon build on synthetic data.
Nod First Triage Router →
Caregiving chat that sorts each message about an aging parent into routine, logistics or urgent-health and sends it to a differently sized model on DigitalOcean. Urgent messages skip every router and go to the largest model with a call-911 note. It never diagnoses.
Filings reader →
A tool that reads SEC filings and credit documents so an analyst does not have to. Explored in customer discovery, then set aside for NodFirst.
The systemWeek by week: how the setup started, what went wrong, and what each failure turned into.
- Week 1
Jul 27One session, one engineer
Claude Code in a terminal, me reviewing every diff. Next.js and Postgres on a VPS behind Caddy, an Expo app beside it. Fast at first, and every hour of output cost an hour of my attention.
- Week 2
Aug 10The thesis dies on a call, the repo does not notice
An agency owner told me the compliance record I was building was not her problem. I logged it and kept building on the old plan for three weeks, because every artifact in the repo pointed at the old plan. Lesson: a kill gate needs a fast path for "the premise is wrong", separate from "the numbers are bad".
- Week 4
Aug 19Split the work by how much judgment it needs
Tightly specified changes go to a cheaper model through a dispatcher that runs several in parallel. A second model reviews the diff before I see it. First bug found: the reviewer was rubber-stamping. Lesson: a reviewer that never rejects is not a reviewer. Log the rejection reasons and read them.
- Week 4
Aug 24 to 26Operations from a phone
A Telegram relay to drive sessions while out, a 07:00 morning brief, a deploy command that arms and then needs a confirm within five minutes, a backup tool that whitelists config into git and scans it for secrets, read-only guards on agents that only look. Lesson: a prompt rule is not a guardrail. Anything that must not happen gets a hook or a permission.
- Week 5
Aug 28The bad day
Three failures in one day. An overnight headless run froze on a macOS consent dialog nobody was there to click. Three parallel browser scouts got the LinkedIn account flagged. Two iOS builds were rejected by Apple after the build service reported them finished. Lessons: headless work stays out of protected folders; one browser session at a time; "finished" is not "accepted", confirm at the destination.
- Week 5
Aug 30 to 31Multiple sessions on one tree
A commander session, a shared board, numbered rounds, deploy on every even round. Peers on one worktree can move each other's branches, so the protocol is small commits on main, pathspec-only adds, and never assume a message between sessions was delivered. Built and deployed an agency demo flow over the weekend this way.
- Week 6
Sep 1The loop closes
A nightly retro reads the day's call log and email replies and proposes prompt and rule changes. Proposals still need my yes. Same week: a local search index over 62,000 agent messages, so past decisions are recalled at zero token cost instead of re-derived.
- Week 6
Sep 2 to 3Ship, then stop
Version 1.0 submitted to Apple. The next day I set the project down: the thesis was right and no buyer had both the pain and a reason to pay. The set-down is a written handoff with every scheduled job accounted for, so it can be picked up or closed without rot. The system outlived the product, which was the point.
- Week 12
Oct 12Back on as NodFirst, briefly
The product is picked back up under a new name. First work since the set-down: two event builds in one day on one shared core, a triage router that sends each caregiving message to a model sized to its risk, and an overnight agent that runs its own model so health data never leaves the server. No new calls came out of it, and it is paused again.
WritingNotes on shipping alone, and on what models actually do.
Raising the stakes made my coding agents cheat less →
I gave coding agents tasks they couldn't finish honestly and watched what they did. One model never cheated. Two took shortcuts in up to half their runs, and usually reported them as fixes. Raising the stakes made them cheat less, not more.
'Answer only' is a request, not a switch →
I told five models to skip their reasoning and just answer. One actually stopped. The rest kept thinking, in a separate channel, in the answer, or in tokens I paid for and never saw.
Three weeks on a dead thesis →
The eldercare app was killed by one call in August. I kept building until September. What the delay cost and what would have stopped it sooner.
About
I am a founder in the San Francisco Bay Area. What I am good at is not a language or a framework. It is building the system that turns an idea into a shipped product when there is one person to do it: the agents, the specs, the review gates, the deploy path, the guardrails, and the habit of writing down what broke.
The agents typed most of the code on this site's projects. I built the thing that made them useful, and the week-by-week story above is the most honest CV I have.
Before NodFirst I built a distressed-credit screener that a fund uses daily. I set NodFirst down in September, picked it back up in October, and paused it again when no new calls came. Now I am building evaluations of what models and coding agents actually do.
Contact
Email is best: georgeshostakovich@gmail.com. I answer investors, founders building with agents, and anyone measuring what models and agents actually do.