Live · 2026
Skophos
AI-curated weekly research digests for biomedical scientists.
3.85M articles · live with INSERM user
Hi, I’m Antoine. I’m 23, French, currently finishing my Master’s between EDHEC and UC Berkeley. I came from finance internships — Crédit Agricole CIB, Amundi, Blue Pearl — and along the way I started building things on the side. I taught myself to code, shipped Skophos to production, and I’m now working on Astra. I want to spend my next years where strategy and execution meet: in early-stage AI startups, as a PM or GTM hire.
I ship AI products end-to-end.
Currently building Astra. On the side, I shipped Skophos (live, used by an INSERM researcher) and Match My Coach. Looking for my first PM or GTM role at an AI-native startup.
Based in Berkeley, CA · Open to Remote / SF / NYC / London / Paris
02 / Work
Three products, taken end-to-end.
From the first user interview to production deploys. Each one started as a real problem and ended with someone using it.
Live · 202601
Skophos
AI-curated weekly research digests for biomedical scientists — live in production, used by an INSERM researcher.
The Problem
Biomedical researchers spend 4–10 hours per week sifting through irrelevant papers. PubMed alerts return hundreds of unranked results. Google Scholar shows what’s popular globally, not what matters to you. Over 2 million papers are published yearly on PubMed, bioRxiv, and medRxiv combined. The signal-to-noise ratio is collapsing.
The Product
A scoring engine combining four independent signals: semantic embedding similarity (Voyage-3-large, 1024 dims), multi-axis LLM evaluation (Claude Haiku, 5 explicit rubric dimensions), citation-weighted bibliometrics (315 journals indexed), and TF-IDF keyword depth matching. Each paper gets a continuous float score [0.00–100.00] — mathematically guaranteed zero ties. Delivered as a weekly email digest of the top 25 papers, plus a web dashboard with archives, discover, and multi-profile research profiles.
The Numbers
- articles in corpus
- 3.85M
- embedded with Voyage-3-large
- 812K
- citation counts backfilled
- 1.77M
- production features in 6 weeks
- 37+
- weekly digests delivered
- 22
- real INSERM user — and counting
- 1
In progress · 202602
Astra
A catalog intelligence layer for the agentic commerce era — the data backbone that keeps brands visible as AI agents take over shopping.
The Idea
AI assistants — ChatGPT, Gemini, Perplexity, Google AI Mode — are starting to route billions of shopping queries. But most product catalogs carry only 5–8 attributes when agents need 30+ to recommend with confidence, so brands risk going invisible in the next era of commerce. Astra connects a Shopify store, auto-enriches every product to 30+ attributes, publishes to the agent protocols (MCP, UCP, ACP), and monitors how the brand surfaces across the major assistants. The data backbone for agentic commerce. First wedge: DTC beauty brands; later, the aggregated cross-merchant dataset becomes an API for AI builders.
Where It Stands
Pre-MVP. Before writing code I produced a 14-section strategic brief — market sizing, competitive map, pricing and a 90-day GTM plan — and I’m now interviewing DTC beauty founders in the Bay Area to find design partners. This is the work I want to do — finding the wedge, sizing the market, designing the pricing, sequencing the GTM — for the right team.
Master Capstone · 2025
03
Match My Coach
A two-sided sports coaching marketplace, built as my 9-month master capstone with a team of five. The product was complete: two-sided onboarding, multi-filter coach search, booking flow, review system, dual dashboards, Postgres + Prisma backend on Vercel. We ran 60 user-research interviews (35 prospective clients, 25 student coaches), designed unit economics, and built a 50-slide investor deck. We never opened it to real coaches and clients — the project ended at the launch gate. The honest learning sits there: product readiness and go-to-market readiness are different problems, solved at different moments.
03 / Background
The path so far.
A triple-campus master between EDHEC Lille, UC Berkeley, and SKK Seoul, with finance and renewable-energy internships in Paris along the way.
Experience
Sales Business Manager Assistant
Crédit Agricole CIB — Global Markets · Paris
Designed and shipped a new onboarding process for booking tools, taking 350+ Sales staff through it. Wrote the bi-weekly markets newsletter for the entire Global Markets department and built VBA dashboards for top management.
Jan – Jul 2025
ESG Risk Analyst
Amundi Investment Solutions · Paris
Translated new ESG regulations into operational controls. Built tools to verify ESG commitment compliance and redefined how Amundi organises its ESG investment constraints across funds.
Jul – Dec 2024
CFO Assistant
Blue Pearl Energy · Paris
Supported M&A in the renewable energy space across France, Belgium and Spain. Sourced and screened acquisition targets; valued companies through teasers, IMs and vendor due diligences.
Jun – Jul 2023
Education
Berkeley Haas Global Access Diploma
UC Berkeley — Haas School of Business · Berkeley, CA
Aug 2025 – Jun 2026
Master in Management — Programme Grande École · GETT track
EDHEC Business School · Lille / Berkeley / Seoul
2022 – 2026
Languages & interests
Languages. French (native), English (C1, TOEIC 965/990), Spanish (intermediate), Chinese (school-level).
Interests. Tennis and rugby (competition), football, sailing, French comics, Asian culture.
04 / Thinking
How I think.
Two lessons from building solo. Not blog posts — things I'd bring into a product discussion.
Why integer scoring breaks recommendation systems
When I shipped the first version of Skophos’s scoring engine, 160 papers came back tied at score 9.0. Top-25 selection became essentially random. I’d assumed integer scoring was good enough — it wasn’t. The fix wasn’t a smarter integer model. It was abandoning integers entirely: continuous floats, aggregated across independent signals with weighted formulas. Zero ties, mathematically guaranteed. I think most recommendation systems eventually converge here. The integer version is the version you ship to learn it doesn’t work.
The feature request is a symptom, not the product
Users ask for features. The instinct is to build them. But the request is almost never the actual product. When Skophos’s first user asked for “more papers per week,” I added a denser digest. He used it less. The real pain wasn’t volume — it was trust in the ranking. The right move was making the scoring transparent (surfacing the 5 axes, explaining “why it matters”), not increasing throughput. I think a lot of product work is this: listen to the feature request, build the underlying job.
05 / Contact
Let’s talk.
I’m looking for my first PM or GTM role at an AI-native startup, starting June 2026. Email is the fastest way to reach me.