Houssem Ben Salem

Houssem Ben Salem, MSc

Founder & architect of Lumin.
I cut the noise and never stop until it's done. I build, I deliver, I lead.
I lead teams of AI agents the way I lead teams of engineers.

2021 — present

Lumin is a systematic directional model for crypto perpetual futures, driven by a cohort of reasoning agents.

+123% on BTCUSDT — February 2025 → May 27, 2026, net of all exchange costs, vs buy-and-hold −12%.
Live through the Lumin terminal since March 2026 — on BTC, ETH, SOL.

The result extends a 6.5-year chained backtest — seven independent cohorts, Sharpe 1.98 net of costs, 7 of 7 years profitable — through the COVID crash, the FTX/LUNA capitulation, the Bitcoin ETF rally, and the Trump tariff cycles.

The architecture

The architecture is a multi-agent LLM cohort: reasoning specialists that read the market bar by bar against a structured curriculum — a practitioner's mind made portable, compounding season over season as the team's curated lessons enter the curriculum, and inheriting the gains of every next-generation reasoning model without retraining. Capital moves only on a verdict the cohort converges on in open debate — and the readers who entered keep the watch until the story ends. The approach generalises anywhere skilled experts produce value.

The origin & the stack

On the Lumin side, the work began from a personal need: cut the emotions out of trades and let an algorithm decide. Years of iteration distilled price action into a structural vocabulary of recurring multi-bar configurations — the seed of the language the current architecture speaks. The production stack is a deliberate two-model design — Claude Opus reads the market; Claude Fable, the newest frontier generation, moderates the debates where capital is decided — with DeepSeek V4 Pro validating the same curriculum across an architecturally different model family. The system extracts edge in choppy markets where deterministic logic breaks down, and captures the asymmetric tail when a clean trend opens.

The harness craft

Harnessing AI agents is a craft in its own right, and Lumin is where I practice it at production grade: fresh zero-memory spawns; blindness by design — no symbol, no calendar, rescaled prices; role isolation declared file by file; a debate moderated relay-only, every cited print verified against the tape. And when a substrate resists — DeepSeek compresses long context where Claude reads it dense — the harness answers structurally: re-inject the framework inside the model's live attention window, or hold the year of bars outside the model as a variable the reader queries through bounded, full-attention sub-reads. Understanding how a model attends is the prerequisite for making it read. The layer above the models is where that edge is engineered.

Two careers, one craft

Lumin advances alongside my main career — fourteen years of Machine Vision engineering at Swiss precision metrology houses. Reading the physics behind a metrology measurement uses the same logic as reading the psychology behind a market price: identify the generator, model its influence, engineer for it. The reading reflexes, modelling tools, and post-mortem discipline that resolve an optical measurement are directly reusable in Lumin. Since 2021 the two trajectories have advanced in parallel, each breakthrough on one side feeding the other.

LUMIN> read the white paper · LUMIN> open terminal
Selected results · net of exchange fees & funding
Annual returns Lumin BTC buy-and-hold
2020+360%+302%
2021+228%+58%
2022 FTX/LUNA+90%−65%
2023+51%+155%
2024+91%+112%
2025+112%−7%
2026 YTD+28%−15%
Compounded total ~224× ~9×

Background
2015 — present

At SYLVAC SA — Swiss precision instruments, Yverdon — I joined in 2015 as an applied optics engineer building vision metrology systems. In 2018 I moved to project management and delivered a generation of optical measurement machines end-to-end: specifications, industrialisation, production. Since 2021 I lead the R&D machines team — still building alongside, pushing the domain into embedded agents, adaptive workflows, machines that interact with humans more naturally.

Watch the testimonial →

Machine vision · optical metrology · quality control · robotic guidance · automation

2012 — 2015

Before SYLVAC, I was an applied optics engineer at Hexagon Manufacturing Intelligence. I worked on metrological machines, optical sensor development, and image processing for precision measurement systems.


Education
2022 — 2024

Applied Data Science: Machine Learning program at the EPFL Extension School — the foundation for the data and Machine Learning work that became Lumin.

2010 — 2012

MSc in Microengineering — Photonics and Applied Optics at École Polytechnique Fédérale de Lausanne (EPFL). Photonics (light guides and lasers), advanced image processing, optical system design.

2007 — 2010

BSc in Microengineering at École Polytechnique Fédérale de Lausanne (EPFL). Optics, vision systems, mechanics, electronics, software engineering.


About me

I work at the intersection of quantitative systems, AI architectures, and precision engineering. My strength is not staying inside one domain but connecting several until they become one solution. I'm an engineer first: every problem reduces to understanding the physics, building the math, writing the code, and delivering it personally — and staying on it longer than most people think is reasonable. I think outside the box, inside the path. Multilingual. I used to code in C++ and MATLAB. Now fully in Python, mostly through AI. English is becoming the new language of code.