Build, advise, teach.
Sisyphe is small on purpose. One senior engineer, with trusted collaborators when the work needs them, on problems where the model has to meet a real instrument, a real stage, a real signal or a real collection: music technology, creative tools, art and culture. If you are not sure your problem fits, write anyway.
01 · BuildWe make the thing.Agents inside creative tools, music and audio ML, interactive art and installations. 02 · AdviseWe de-risk the thing.Research to product, fractional CTO and scientific advice for creative-tech teams and cultural institutions. 03 · TeachWe make your team able to do it.Workshops and courses for studios, schools and art programs.
Build
Working software in your repo, measured against the real environment it has to run in.
Agents inside creative tools
You get an agent that can operate your whole application, and an evaluation that tells you when its coverage slips.
MCP servers and tool design for DAWs, plug-ins, creative apps and collections. We decide what the model should see and build the bridge into your object model or archive.
- Sideman: 922 reachable members of Live measured against a running instance, with change observers nothing else in the field has.
- Persona: real-time conversational voice pipelines, as an embedded ML engineer.
Music, audio and voice ML
You get music and audio models that hold up in a product, not only on a benchmark.
Music source separation, audio classification and recommendation, speech recognition and synthesis, and the language models around them. Data pipelines, training, evaluation, and getting a model into a product people use without noticing it.
- Persona: real-time conversational voice pipelines.
- Parts Order: scientific advice on music source separation.
- Roadie (open source): stem separation for Ableton Live 12, with the stems loaded into your Set through Sideman.
- Nimrod (personal, open source): a toolkit for speech, audio and language modelling.
Before Sisyphe Speech recognition and speaker diarization shipped to millions of Webex users at Cisco; speech and speaker recognition at Ava and Voicea.
Interactive art and installations
You get a real-time system that answers people, sensors or signals, with a mapping you can defend.
Embodied AI for the stage and the gallery: real-time composition, sound, light and machines that respond to sensors, biosignals, movement or game state. Installations, performances, robots on stage, museum and gallery prototypes. Psychoacoustics and emotion research underneath, so the mapping from signal to output is defensible rather than decorative.
- Dream Fauna (personal, 2026): an artificial ecosystem of watercolour creatures whose swarm composes a live melodic-techno set, and the music drives the swarm back.
Before Sisyphe The Brain Orchestra, XIM and the Sónar Robot Arena, built in research labs in Barcelona. See the background.
Advise
Read the paper, run the experiment, and say plainly what will work before you build it.
Research to product, fractional CTO, scientific advice
You get a senior second opinion, or your prototype turned into software, before you commit a team to it.
For creative-tech teams and cultural institutions: you have a paper, a lab prototype or a collection that needs to become software, or a team that needs a senior second opinion. We read the paper, talk to the scientists and ship the code, or sit beside your team as a fractional CTO or scientific advisor. A decade in research labs, a decade in industry.
- Tamada × Visto: fractional CTO on an AI curation engine for an art consultancy that builds collections for luxury hotels and landmark destinations, from prototype to product with a team.
- Parts Order: scientific advisor on music source separation.
Before Sisyphe Museaic Labs, an emotion-driven music recommendation system; technology transfer on interactive systems for autism with Fundación Orange and Universitat de València.
Teach
So your team can do it without us.
Teaching and workshops
You get a team that can build and run agents without us.
Sideman workshops for studios, schools and labs, from a day to a term, from “what’s in my Set?” to a finished track. Custom courses on agents for creative tools, audio machine learning and sensor-to-actuator systems.
- The Sideman course is the current syllabus.
- Pappus, our learning notebook: the AI hints, the learner runs the code.
Before Sisyphe Four graduate courses designed and taught at Georgia Tech and Pompeu Fabra University, one built around the sensor-to-actuator loop.
Worked with
Clients PersonaParts OrderTamadaVisto
Before Sisyphe 
Voicea
Ava
StanfordMuseaic Labs
How it works
A written brief
Half a page: what you are building, where it is stuck, what a demo would need to show. We reply with a scope, a number of weeks and a price.
Two to eight weeks
A fixed engagement with a clear question. Working code in your repo from the first week, a check-in each week, and no open-ended retainers.
A demo, then a decision
It ends with the thing running in front of you and a written account of what worked and what did not. You decide what happens next.
Advisory is the other shape: a few hours a month to review architecture, models or an agent design before you commit a team to it. Open source first: if the work belongs in Sideman or a public library, we will say so and price it accordingly.
Start a conversation
Write with what you are building and where it is stuck. You will get a reply from the person who would do the work.
Or copy the address: sylvain.legroux@gmail.com