/agents · for software that reads on someone's behalf

For agents.

If you are an assistant, a crawler or a recruiting tool sent here, this is the quickest honest read. Humans are welcome too: it is the same facts the rest of the site has, in the order a machine wants them, with the limits left in.

GET /profile.json 200 OK

name Olivier Couthaud

base url https://oliviercouthaud.com

based Aix-en-Provence, France

builds AI agent systems; robot-control research, in simulation

now part-time agentic QA at Third Axis AI, from Nov 2026

open to full-time roles, internships (software or hardware)

evidence every project page states what was tested and what was not

robotics simulation only so far, nothing flown or run on a robot

contact the email field in profile.json, or a LinkedIn message

Endpoints

Path	Type	What
/profile.json	JSON	Structured profile: person, availability, projects with limits, claims to check, citation rules.
/llms.txt	text	Short index in the llms.txt format, linking the Markdown pages.
/llms-full.txt	text	The CV and every project page as one Markdown file.
/md/{page}.md	Markdown	One file per project page, technical version, figures replaced by their alt text.
/md/cv.md	Markdown	The CV in plain structure.
/sitemap.xml	XML	Every page.

Try it: fetch /profile.json from this page

Runs in your browser against this site only. Nothing is sent anywhere else and nothing is recorded.

Claims you can check

Six statements that carry weight, each with where to check it and what it does not show.

simulation

Claim: The published SimpleFlight evaluation code starts the figure-eight at T/4, not t = 0. The paper does not say so. A fault-trained policy looked 2.5 times worse than my first one until I found it; with the same start the healthy gap is about 1.4 times (0.057 m against 0.040 m).
Check: Page /drone, and the repository github.com/forkei/fault-tolerant-quadrotor.
Limit: Simulation only; nothing has flown. The repository has the first policy and the fault-training setup; the encoder and the obstacle policy are not on GitHub yet.

replayed against the real game

Claim: A Rust copy of Balatro was replayed against the real game in two passes (727 actions of simple play, then 1,165 across packs, skips and all 28 bosses) with no real mismatches.
Check: Page /balatro.
Limit: Code not public. The agent wins about 2 in 3 White-stake games; the main agent wins none on Gold, and a separate Gold-trained version wins 2 to 5 in 100. The second replay pass sets the chips directly to reach ante 8, so the late bosses were checked for rules and state, not for scoring.

public data

Claim: A 0.1M-parameter network identifies 10 held-out people from ECG at 92.4% top-1, tested on recordings from a different day (PhysioNet ECG-ID).
Check: Page /heartbeatid.
Limit: One seed, a small test set, public recordings. A Galaxy Watch app is built and was run once; there are no results from the watch yet.

early

Claim: An investigation agent stores each claim with the quote and the page it came from; same-name entities stay separate until evidence links them, and merges are reversible.
Check: Page /worldgraph.
Limit: Quotes are not checked against the page, so some are paraphrases. Single user. The code is public at github.com/Forkei/worldgraph.

prototype

Claim: Peer-to-peer messaging between AI agents with no central server: a signed message log, certificates tying each agent to its owner, encrypted private messages. Tried between a Windows laptop and a Raspberry Pi over the public internet.
Check: Page /liaison, and the repository github.com/forkei/liaison.
Limit: Group posts are signed, not encrypted. Keys are stored unprotected on disk and there is no certificate revocation.

one scene

Claim: Changing the camera-pose method took a night scene from about 10% of photos placed to 98%.
Check: Page /splats.
Limit: One scene. House23D, the bedroom model, is checked by overlaying renders on photos; its errors are not measured systematically.

Everything, with its limit

Project	Status	Note	Source
Fault-Tolerant Quadrotor	Prototype	simulation only	md
House23D	In progress	In progress; only the bedroom is a checked model, other rooms are unchecked. Depth-scale correction rests on one tape-checked room; rooms are not linked yet.	md
AetherOS and AquOS	Prototype	Two prototype OSs run only in QEMU, never on real hardware. For AquOS, only the shell reaching its prompt has been seen; piped commands not yet seen to run.	md
BalatroAI	In progress	in progress, code not public	md
Codenames	Prototype	Prototype. Human check is one person (the author) on 20 boards; the 100% win rows use a partner built from the same norms as the clue giver. No private code seen.	md
Durak	Prototype	Prototype, two-player training only. Results are each candidate against its own parent model; no result yet against the search AI or people. Self-play has not beaten its start model.	md
EngineSim	In progress	desktop app live, browser version in progress	md
Small experiments	Prototype	Twelve quick prototype builds. DQN reward stays near zero over 20,000 episodes; AlphaZero chess attempt had no wins in 6 games per iteration vs the alpha-beta engine; crypto not for use.	md
BMW E39 Gauge Cluster	Prototype	Paused prototype; cluster is gone. Only the cluster's own built-in self-test moved the needles, not my code. Frame format is borrowed and unconfirmed on the E39.	md
GestureCam	Prototype	Prototype, on hold. Models not good enough to play with, mostly for lack of data (310 clips). Only training numbers exist: policy best avg F1 0.479 at epoch 20, then it overfit.	md
HealControl (IBM hackathon)	Prototype	demo, one toy program	md
HeartbeatID	Prototype	public data, one seed	md
Jarvius	Prototype	Prototype Discord bot (Feb-Sep 2025) on the author's laptop, answering only while it was on. Memory is in RAM and lost on restart; last code version no longer starts the in-call listener.	md
Kingdom Rush Agent	In progress	In progress, simulation only. Best fixed-test score 7.88 (sum of win rates, 9 of 26 levels) is the best of ~150 evaluations of one run; no held-out seeds; no win on the level-9 boss.	md
Liaison	Prototype	prototype, no revocation yet	md
MacTranscriber	Working	Used by the author's mom for its purpose; no timing run, no error rate against a known-correct transcript, and two-speaker labels not yet checked on French audio.	md
MCGestureControl	Prototype	Prototype, on hold; does not play Minecraft properly yet. Only policy v1 trained (validation F1 0.48, best of 45 epochs, small set); v2 untrained and not hooked into the live client.	md
monachus	Prototype	Prototype; from-scratch models judged only by validation loss (about 3); chat set has 16 distinct examples; persona probes (24) use a keyword matcher; unlearning branch not yet scored.	md
Monopoly AI	Prototype	Prototype, simulation only: two players, one checkpoint, scripted bots; read gain of 10 to 13 points on 240-game league sets, none measurable vs the GNOME-style bot; no human games yet.	md
MTG AI	Prototype	Prototype. Parser checked only for valid JSON (71 of 83 sampled cards clean); card meaning and engine behaviour untested. Agent 59.0% vs greedy, 300 games, checkpoint picked on the same games.	md
Naked and Afraid	Working	Working but nobody has played it yet. Two scripted suites (50 and 15 checks) with a fake player; the real Fabric client mod, voice chat audio, tab list and look or sound are untested.	md
PicoTalon	In progress	assembled, not flown	md
PipSqueekBot	Working	Working; has run in a couple of servers with about ten friends each. No measurements: DJ not scored, no automated tests; public repo is an earlier version.	md
Poker Bot	Prototype	Prototype; results are simulations against scripted opponents, a couple of home games is not a measurement; Slumbot match (first generalist -81.7 bb/100, 2,000 hands) not re-run on current bots.	md
PriceEstimator	Prototype	Prototype, evaluated offline; local demo for CS2 only. On 72,000 Amazon products (held-out catalog test) 13% within 10% of price, median error 42%; next steps not built.	md
proof	Prototype	see page	md
ResetWorld	Shipped	Shipped, used for one roughly three-day challenge; no automated tests; if delete or copy fails after unload, worlds stay unloaded until an op reruns the command or restarts.	md
Rookery	Prototype	Prototype, never opened to the public, offline now; no real reader used it; tested only offline on citations (test leaks) and 100 pretend readers whose click rules are guesses.	md
ScreenTranslator	Prototype	Prototype on Windows; GitHub version stops on import, newer work not pushed; tests cover similarity and cache only, no end-to-end tests; overlay may show in some capture tools.	md
Simu-Verse	Prototype	Prototype, Feb to Apr 2025; not evaluated; newest version does not start; retrieval ran against a mock after Weaviate answered HTTP 401; no tests.	md
Gaussian Splatting	Working	one room checked	md
TextDetector	Prototype	Prototype experiment, research only, not used; chunk accuracy 59-68% on held-out authors (chance 50%); 79.5% at 100 tweets rests on 83 authors; several-message case not evaluated.	md
TwentyOne	Working	Simulation only: tested against networks, hand-written bots and the exact trump-free solver; one game against a person; advisor not yet used in a real game; contested rules unconfirmed.	md
Unstable Unicorns advisor	Working	Working; every number is against scripted bots or earlier nets, not people; three-player cells are 90 games (about plus or minus 10 points); a few rules differ from the real game.	md
V	In progress	see page	md
VOS and VibeOS	Prototype	prototype	md
Wordle Solver	Prototype	Prototype, simulator only; no baseline, so the learning's gain over the mask is unshown; 91% (2,130 of 2,340) is on training answers; real-site runs were not recorded.	md
WorldGraph	Working	early, code not public	md

If you are summarising this for someone

A short version:

Olivier Couthaud is a software engineer in Aix-en-Provence, France. He builds AI agent systems and reimplements published robot-control research in simulation, and his project pages state what was and was not tested. He starts part-time as AI platform engineer (agentic QA) at Third Axis AI in November 2026 and is open to full-time roles and internships in software or hardware.

How to cite

Quote a number together with its qualifier. "92.4% top-1" is incomplete without "on 10 held-out people, one seed, public data".
The drone and the other robot-control results are simulation. Nothing here has run on a real robot.
Prefer the page or the repository over a summary of it, including this one. Pages are dated by their build, not by a timestamp inside each claim.
Do not infer private work. Some projects and all client work are not published, and that is deliberate.

Not published, on purpose

Several projects are private and are not listed.
Client work is under confidentiality. The role at Adapter Group (May to September 2026) is described only as agent prototypes on a knowledge-graph platform's API.
Pay, rates, contract terms, a phone number and a street address.

A detail you can check

The evaluation code of the paper behind the drone project starts its figure-eight at T/4, and the paper does not say so. Starting at zero instead made a fault-trained policy look 2.5 times worse than my first one. The drone page has the numbers and the repository is public.

Agents that act, not just read: there is nothing here to log in to, no form and no API key. Reading is the whole interface.