Jordan Levy

Data scientist at Intel. The rest of the time I build agents, benchmarks and games.

At work I help build a retrieval-augmented assistant and the evaluation tooling that checks it. At home I run coding agents on three machines and use them to ship the projects on this page. Most of them you can open and try.

See the work Play the games

At Intel

Data scientist since 2020. Most of it is internal, so this is the short version.

  • An assistant that knows where to look Helped design an enterprise retrieval-augmented assistant that sends each question to SQL telemetry or to technical documentation, depending on which one can answer it.
  • Evaluation around it Built the tooling that checks the assistant: automated checks, batch tests, human review and prompt tuning, aimed at hallucinations and wrong tool choices.
  • Workflows for the team Wrote more than 30 reusable AI-assisted development workflows, each with a test harness and automated checks, for a team of about 20.
  • A 12x speedup Used cProfile and Polars to help make an internal data-processing and performance-analysis platform 12 times faster.
  • Telemetry pipelines Build and maintain pipelines that process several terabytes a week in Python, SQL and PySpark.

Agents and the tools around them

Side projects that each started as a question about what an agent can be trusted to do. The numbers come from the repos and the run logs.

62 model and council setups tested. None beat the index by more than luck.

CouncilBench

Runs weekly

A benchmark that asks whether LLMs, alone or voting as a council, can beat an index fund after costs and taxes by more than luck would explain. The test was preregistered and then frozen. Every decision goes into a tamper-evident journal, and a masking audit checks whether a model is just remembering the dates.

Setups scored
62
Paper portfolios
7, at $10,000 each
Test files
76

Python, uv, Claude, Codex and Ollama model runners, Alpaca paper trading

Baseline's Today page showing sleep, heart rate and activity against a usual range. The data in this screenshot is invented.

Baseline

Live

A health dashboard for one person. It merges Apple Health, Google Health and a RingConn ring, then compares each day with my own usual range. Agents read it through an MCP endpoint with seven read-only tools. Each agent gets its own scoped token that I can revoke, and every read is logged where I can see it.

Agent access
7 read-only MCP tools
Tokens
One per agent, revocable
Hosted
On this server, behind a login

TypeScript, Vite, Model Context Protocol, Playwright, nginx and systemd

One run reached the Act 1 boss in 175 actions, for four cents.

Slay the Spire 2 Bench

Results live

A local harness that measures how far an agent can get in Slay the Spire 2 through a headless command line. Runs are seeded and traced action by action, the game runs inside a macOS sandbox, and a dashboard replays saved or live traces.

Best run so far
Act 1 boss, 175 actions
Cost of that run
134,614 tokens, $0.039
Codex threads
198
Published runs
40, across 9 models

Python, .NET, OpenAI and OpenRouter

See the results
It suggests. I decide.

FFAgent

Runs in season

Fantasy football for my Yahoo leagues. It pulls projections from several sources, writes waiver reports and a weekly outlook, and runs on a schedule on the Mac Mini. It only makes suggestions. Every roster move is still mine.

Reports
Tue, Wed, Thu and Sun
Commits
52

Python 3.13, PostgreSQL, OpenAI, Anthropic or local models

Also running

  • Larry's Gym Tracker A Discord bot my friends and I use to log workouts and sleep, with a points game on top. It has been running since December 2023.
  • Music Decomposer Turns a song into a guessing-game video. It decomposes the track with an FFT, renders the video with FFmpeg and writes the upload metadata.
  • Basin Bike routes around Los Angeles and Orange County that change based on what I say about each ride. Self-hosted on the Mac Mini so my agents can reach it.
  • DFS Lineup Lab Builds NBA daily-fantasy lineups every day and scores them on paper against real contests, so I can see whether the model is any good before it matters.
  • PiCar agent A voice agent on a small robot car: offline speech recognition, an LLM for the thinking, and a synthetic voice for the reply.
  • Gas bill splitter Reads the gas bill from email, splits it between roommates and files the payment requests. Runs on GitHub Actions.
  • Streetball Revival Vj's from-scratch Rust rebuild of a 2003 PS2 streetball game, with online play. I contribute to it. 42 of the commits so far are mine.

Games

Simulations, mostly. Games are where I practice letting agents do real work, because a game either plays well or it doesn't.

Lemonade Stand mid-day, with customers walking past the stand and the day's sales counted below.

Lemonade Stand

Live

A roguelike business sim. You run the stand during the day, then draft a card each night that bends the rules of the run. Two people can share a market.

Play in your browser How it was built
The AGI Tycoon dashboard showing daily demand, revenue and a market share chart for five competing labs.

AGI Tycoon

Live

You run an AI lab against four simulated rivals. Train models, price them, raise money and handle alignment incidents while the market moves under you.

Play in your browser How it was built

Smaller ones

  • Math Tug-of-War A one-on-one math duel on two phones over the local network. React, Express and Socket.io.
  • SuikaBot A reinforcement-learning agent that learns to play Suika from screenshots.
  • Wiki-Drop Guess where a historical event happened from its Wikipedia summary, with the giveaways scrubbed out by a model first.
  • Shadow Boxr map generator Python tooling that turns songs into boxing rhythm maps in the Beat Saber format.

How I work with agents

I use several coding agents every day and have opinions about each. This is the setup and the rules that survived.

There are three machines: a MacBook Pro, a PC, and an M4 Mac Mini that never sleeps. The Mini is the home server. It runs about fifteen of my own background services, including the fantasy football reports, two Discord bots, a couple of agent gateways, local models through Ollama and LM Studio, and a self-hosted CI runner.

The laptop is where I sit. Planning and review happen with one model, bounded implementation goes to another, and parallel work happens in separate git worktrees so two agents never touch the same files.

What I use, and how much
Tool Volume What for
Codex 1,870 threads Since September 2025 on the laptop alone. 634 of them were on AGI Tycoon.
T3 Code 146 threads Six projects on the laptop, and a second T3 Code server on the Mac Mini.
Claude Code Daily Planning, review and anything a person will look at. It hands bounded implementation work to Codex.
Scheduled agents 13 automations Five running now, including fantasy football reports four days a week.

  • Tests come first, and the failure has to be seen

    My rules file tells every agent to write the test, watch it fail, then write the code. An agent that skips the red step tends to write tests that cannot fail.

  • Route work by what it needs

    I keep a table that scores each model on cost, intelligence and taste. Architecture and anything user-facing goes to the model with the best judgment. Bounded refactors go to the cheapest one that can do them.

  • One implementer, at most one reviewer

    That is the cap in the AGI Tycoon rules. Parallel work happens in separate git worktrees, on separate files.

  • Give the agent a way to check itself

    AGI Tycoon has a headless autoplay harness and a telemetry catalog, so a balance change is judged on recorded runs. The agent reads the evidence the same way I do.

  • Nothing ships without me

    Agents do not push or deploy on their own. CouncilBench runs under spending caps. FFAgent only makes suggestions. The health dashboard gives agents read-only tools and logs every read.

  • Write the goal down

    Long tasks start from a goal file and a spec, and keep a progress log. When a session runs out of context the next one picks up from the files.

About

I studied data science at UC San Diego, with a minor in cognitive science, and joined Intel as an intern in 2020. I have been there since.

Most of what is on this site started as a question. Can a committee of models pick stocks? Not so far. Can one get through a deck-builder? To the first boss, for four cents. Can I leave one running on a schedule and trust it? Yes, if all it can do is suggest.

More about me