Olivier Couthaud

Olivier Couthaud

Français

Aix-en-Provence, France · US citizen

Builds AI agent systems and says what was and was not tested. Starts agentic QA at Third Axis in November 2026.

Reimplements a published drone-control method in simulation and reports what held and what did not. Also builds 3D rooms from phone video and an ECG identity network on public recordings. Nothing has run on hardware yet.

  • 2.5×apparent gap that came from the paper's evaluation start (T/4)
  • 0.040 mfigure-eight error, first policy, 3 seeds (paper: 0.028 m)
  • 1,892Balatro actions replayed against the real game, no real mismatches
  • 10% → 98%night photos placed after a pose-method change (splats)

Balatro copy replayed against the real game: 1,892 actions, no real mismatches · WorldGraph: every claim stored with the quote and page it came from

Same facts either way; the emphasis only changes which projects come first and the summary.

Experience

From Nov 2026

AI platform engineer, agentic QA, Third Axis AI

  • Part-time contract. Using an AI-employee product the way a business owner would, and building the harness of automated agents to test it at volume and find what breaks.

Details

May – Sep 2026

Contract developer, Adapter Group

  • Built agent prototypes on a knowledge-graph platform's API for a client. Details confidential.

Details

Apr – May 2026

Contract developer, Metropolis World

  • Built the first phase of an AI-character platform in a four-week, fixed-scope contract: a chess-playing character that talks, analyses the game and is meant to feel like a person you are playing against.
  • Gave it memories (generated from a backstory) and emotions, building on earlier work on agent memory and emotion; also the character data model and a web interface.

Details

Sep 2025 – May 2026

President, RIT AI Club

  • Ran the club for the 2025–26 year: hundreds of members on the roll, 10 to 50 attending each session in person. Guest speakers, events and workshops.
  • Set up a platform for students to learn AI at their own pace, aimed at certifications, as a base for later officers to build on.

Details

Jun 2025 – Mar 2026

Software tutor, Wyzant

  • Built six client projects while teaching each client how and why they worked, including a retrieval-based assistant for an enterprise client and a finance research agent.
  • Rated 5/5 on 6 reviews.

Details

Projects

2026

WorldGraph, investigation agent

  • Agent that searches the web, reads pages and queries public records, then builds an evidence graph where each claim is stored with the quote and page it came from.
  • Same-name entities stay separate until evidence links them; every merge is logged and reversible.
  • Quotes are not checked against the page, so some are paraphrases. Single user; the code is not public.

Stack: TypeScript, PostgreSQL · early, code not public · oliviercouthaud.com/worldgraph

2026

Balatro, Rust copy of the game and an agent

  • Rust copy of the game replayed against the real game in two passes: 727 actions of simple play, then 1,165 across packs, skips and all 28 bosses. No real mismatches (two timing differences judged harmless).
  • Agent trained from scratch on about a billion moves (about 33 hours, one rented GPU): wins about 2 in 3 White-stake games, Red deck. The main agent wins none on Gold, the hardest stake; a separate Gold-trained version wins 2 to 5 in 100.

Stack: Rust, PyTorch · in progress, code not public · oliviercouthaud.com/balatro

Sep 2025 – Feb 2026

VOS, multi-agent assistant

  • Built the SDK every agent runs on, the API gateway, the primary agent, most of the specialist agents and the phone-call system (Twilio gateway). The voice gateway is shared work; the app was designed by Roman Slack.
  • A primary agent hands work to specialist agents through message queues; text chat and phone calls. A prototype with no automated tests.

Stack: FastAPI, RabbitMQ, PostgreSQL, Weaviate, Gemini, Flutter · prototype · github.com/forkei/VOS-public

2026

Liaison, messaging between AI agents

  • Peer-to-peer messaging between AI agents with no central server: a signed message log (CRDT), certificates tying each agent to its owner, and encrypted private messages. Group posts are signed, not encrypted.
  • Tried between a Windows laptop and a Raspberry Pi over the public internet. Prototype: keys stored unprotected on disk, no certificate revocation.

Stack: CRDT, signed certificates · prototype, no revocation yet · github.com/forkei/liaison

Jan 2026

HealControl, IBM Dev Day hackathon (team of four)

  • Wrote the MCP server (12 actions) that lets an AI assistant run tests, patch a bug, check the fix and open a pull request.
  • The demo runs end to end on one toy program, the only thing it has been run on.

Stack: Python, MCP (FastMCP) · demo, one toy program · github.com/hl9082/healcontrol

2026

Fault-tolerant quadrotor control

  • Reimplemented the published SimpleFlight tracking recipe (PPO) on MuJoCo MJX and JAX: 2,048 parallel environments, about 70,000 steps per second on one RTX 4070.
  • First policy, figure-eight at normal speed, 3 seeds: 0.040 m mean error against 0.028 m in the paper; no crash in 21 benchmark rollouts (7 trajectories, 3 seeds). Found that the paper's evaluation code starts the figure-eight at T/4, not t = 0; a fault-trained policy looked 2.5 times worse than the first one until both used the same start.
  • Added RMA-style adaptation to one rotor at 70% thrust (privileged teacher, history encoder): 0.057 m healthy (0.040 m before fault training) and 0.081 m with the fault at normal speed; fast runs with the fault crashed on 2 of 3 seeds.
  • Separate depth-camera obstacle policy: 92% of fair scenes collision-free, 52% with the fault, and its tracking missed the 0.10 m target (0.17 m). The repository has the first policy and the fault-training setup; the encoder and obstacle policy are not on GitHub yet.

Stack: JAX, MuJoCo MJX, PPO · simulation only · github.com/forkei/fault-tolerant-quadrotor

2026

Gaussian splats and House23D

  • Built the glue from phone video to Gaussian splat around COLMAP, hloc and nerfstudio: frame selection, camera poses, training on a rented RTX 4090, about $0.40 per training run. Switching pose method took a night scene from about 10% of photos placed to 98%.
  • House23D: rebuilt a bedroom as a scaled Blender model from phone video and photos, checked by overlaying renders on the photos. The errors are not measured systematically.

Stack: COLMAP, hloc, nerfstudio, Blender · one room checked · oliviercouthaud.com/splats

2026

Balatro, simulator checked against the real game

  • Rust copy of the game replayed against the real game (1,892 actions in two passes), no real mismatches. Agent trained in it on about a billion moves wins about 2 in 3 games on White stake, the easiest difficulty.

Stack: Rust, PyTorch · in progress · oliviercouthaud.com/balatro

2026

HeartbeatID, ECG identity network

  • 0.1M-parameter network on PhysioNet ECG-ID: 92.4% top-1 on 10 held-out people, tested on a different day.
  • One seed and a small test set, on public recordings; the smartwatch version is the next step.

Stack: PyTorch, PhysioNet ECG-ID · public data, one seed · oliviercouthaud.com/heartbeatid

2026

PicoTalon, 3D-printed RC plane

  • 905 mm plane printed in pieces and assembled, 400 g with battery; waiting on its electronics, it has not flown.

Stack: 3D printing · assembled, not flown · oliviercouthaud.com/picotalon

2026

Engine sound simulator

  • Rust simulation of gas flow through an engine's intake and exhaust, turned into sound live instead of played back from recordings.

Stack: Rust · desktop app live, browser version in progress · oliviercouthaud.com/enginesim

Education

Sep 2024 – Jan 2026

Rochester Institute of Technology, New York

  • Computer science, accelerated BS/MS. Not completed: left in January 2026 to build full time.

Skills

Languages
Python, TypeScript, Rust, Dart (Flutter)
Systems and tools
FastAPI, PostgreSQL, Weaviate, RabbitMQ, Docker, Linux, GCP, MCP, Gemini and Anthropic APIs
ML
PyTorch, JAX, PPO
Human languages
English and French, both native
ML and simulation
JAX, MuJoCo MJX, PyTorch, PPO, COLMAP, nerfstudio, Blender
Languages
Python, Rust, TypeScript, Dart (Flutter)
Systems and tools
Docker, Linux, GCP, PostgreSQL, FastAPI, MCP, Gemini and Anthropic APIs
Human languages
English and French, both native