English 
한국어 
Français Own product · in development
Where?
344 designed flow steps
Plan outings alone or in groups, then reveal the map by actually going. Product, UX, iOS app and backend — all mine, built with a team of AI agents.
Read the case studyR&D case study
MoEVM — a 58 GiB model on a 48 GB Mac.
Mixture-of-Experts models only use a few experts per token. MoEVM keeps the right ones on the GPU and pages the rest from the SSD, so models beyond the GPU’s memory budget run on Apple Silicon — with output identical to llama.cpp without MoEVM.
- 8.72 tok/s gpt-oss-120b (58 GiB) on the GPU, median of 5
- 3.1× Qwen3-Coder-Next vs llama.cpp op-offload, same prompt
- 0.0 max logit difference — 1,024 rows × 151,936 tokens
- 95.8 % expert-cache hits with LRU (optimal: 97.5%)
- 1 Pack
Routed experts packed into a sidecar file — 4,608 records with a SHA-256 index.
- 2 Stage
A RAM arena with LRU eviction; 4 threads prefetch the experts the router is likely to pick.
- 3 Route
Read the router’s choice once per layer, plan GPU slots, swap expert ids atomically.
- 4 Compute
Experts run from a persistent Metal pool; the rest of the model stays on the GPU.
- 5 Protect
Epochs: a slot is never evicted before the GPU is done with it. A memory-pressure listener trims RAM.
- 6 Verify
Every change is gated by logit parity against llama.cpp without MoEVM.
- llama.cpp op-offload 4.24
- Sidecar + Metal pool 13.08
- Hardened, dynamic-K 19.31
- 4k context 21.88
- Graph-build, no copy hook 21.90
Qwen3-Coder-Next Q4, tokens/s, runs with routing traces on. Clean head-to-head with tracing off: 6.41 → 19.90 tok/s (3.1×).
How I kept it honest
- Parity row by row: max difference, argmax and token over the full vocabulary.
- Jitter test: 500 µs injected delays and forced syncs — still 0.0.
- Launch gates written before the results; median of 5 runs, never the best one; warm-cache numbers refused.
What didn’t work — and the limits
- Callback-based expert loading: 2.07 tok/s. Dropped.
- When a model fits in memory, plain llama.cpp wins (49.69 tok/s). MoEVM is for the ones that don’t.
- On CPU, plain llama.cpp runs gpt-oss-120b at 10.9–14.4 tok/s on short prompts. MoEVM’s edge is the GPU path with bounded memory.
- Speculative decoding brought no real gain on the final path.
More case studies
Founder & CEO
Selvren
11,901 tests in the merge gate
AI domain experts grounded in a company’s private documents — every claim checked against its sources before it’s shown. Dashboard and enterprise API.
Read the case studyLet’s build something.
I take freelance projects: I design, build and ship AI tools wired into the software your team already uses. Also open to joining a team or a startup that needs someone who builds — remote-first today, regular stays in Seoul, and a move to Korea ahead.
Or reach me directly
- Email maxence@selvren.com
- WhatsApp +33 6 19 32 62 26 Open ↗
- KakaoTalk ID Maxence.foulon
SouverainIA
AI executive assistant inside WhatsApp, hosted in France.
Context
SouverainIA’s founder wanted an AI executive assistant for SME leaders that lives where they already work — WhatsApp — with every byte hosted in France.
What I did
- Built the product solo: Python backend, WhatsApp (Meta) integration, a “Chef IA” that sets daily priorities and sends nightly summaries.
- Security: TOTP two-factor auth, session revocation, privacy-safe logs.
- Operations: Docker on a self-managed VPS behind nginx, versioned releases with a written rollback for each deploy.
- Designed and shipped the new landing page.
Result
Live at souverainiafr.com, covered by 1,749 automated tests.
Bastide
Private AI assistants for SMEs — offer, pricing and site.
Context
With a partner, a studio that builds private AI assistants for SMEs and firms handling confidential data — accountants first.
What I did
- Positioning, offer and pricing: paid audit, 6-week build, optional monthly follow-up.
- Brand and the full site: Next.js static export, industry pages, a “Bastide vs public AI” comparison, a security page.
- An SEO + AI-search content loop: scheduled drafts, competitor watch and a monthly visibility report — nothing publishes without human review.
ESLITECH
Manufacturer site redesign on their live CMS.
Context
ESLITECH (brand Smartfurry), an aquarium and pet-products manufacturer, had a dated site. The brief: same content and pages, a premium look.
What I did
- Rebuilt the front end in Vue 3 + Vite on their live CMS API, keeping every URL, call-to-action and popup working.
- A motion system, bento product sections and a dedicated About page.
- Shipped a public demo on Cloudflare Workers: HTTPS, noindex, forms disabled.
Selvren
AI experts on private documents, every claim checked against its sources.
Context
Teams want answers from their own documents — with proof. Selvren started as a chatbot for pet shops, became AquaChat for aquarium stores, then a B2B platform of AI domain experts.
What I did
- Founded the company and led the build with a team of AI coding agents: I set the architecture, the rules and the merge gate, and reviewed the work.
- Hybrid retrieval over private documents and an expert catalogue, then a rule-based verifier that checks every claim against its sources — it abstains or quotes when sources are thin.
- Security by design: row-level security in PostgreSQL, encrypted originals, antivirus on uploads, an egress proxy, hashed and revocable API keys.
- GDPR tooling: data export, erasure, a written retention doctrine.
Result
3,182 commits in 11 weeks, 11,901 tests in the merge gate, 28 architecture decision records.
Atelier
Raw idea → product brief through 8 specialist engines, 100% local.
Context
Every product I build starts as a raw idea. Atelier turns it into a brief an engineering team — human or AI — can execute.
What I did
- A local-first chain of 8 specialist engines, from discovery to product, with every step’s data persisted.
- Runs entirely on a MacBook (M5 Pro) with Ollama and MLX, under a strict zero-cost rule: no paid API, no rented compute.
- Benchmarked local coding models to pick the engine (qwen3-coder 30B: 5× faster than qwen2.5-coder 32B).
Result
Idea → product brief end to end in about five minutes, fully offline.
Naiad
Connected-aquarium companion app with an AI assistant.
Context
Naiad is a companion app for a hardware partner’s connected aquariums — built as a partner demo.
What I did
- Designed the whole flow: account-first onboarding with EU data residency, product recognition by QR code, tank profile.
- A living dashboard (water parameters, a confidence score), species compatibility, device control for lights and pumps, a logbook.
- An AI assistant that explains trends and prepares safe actions — but never controls a device without the owner’s confirmation.
GLPI
Set up GLPI for the service desk and IT inventory.
Context
The IT department supports staff across a whole university campus — every request and every piece of equipment has to be tracked.
What I did
- Put GLPI in place: a ticketing service desk for user requests and an inventory of the IT equipment.
- Used it day to day with the team for support follow-up.
PC deployment
A deployment server to install staff workstations.
Context
New and reinstalled workstations had to reach staff ready to use, the same way every time.
What I did
- Set up a deployment server to install work PCs in a repeatable way.
- Prepared and installed staff workstations.
Support & maintenance
User support, network and server maintenance.
Context
Two years as an apprentice inside the IT team of a university, alongside a Bac+2 in systems and networks.
What I did
- Day-to-day support for university staff.
- Maintenance of the network and the servers with the team.
- Hardware maintenance of workstations.
DataRace
A 223-mod “Create Civilisations” server, tuned and hosted.
Context
Still building for Minecraft: a modded survival server built on top of the “All of Create” pack.
What I did
- Assembled and shipped seven versions in five days (June 2026): colonies, world generation, trains, far-view rendering, a performance pass, then content and quality-of-life mods.
- Wrote down every choice: no land-claim mods, hard mob packs blocked, each mod tested on a staging world before going public.
- Tuned and hosted it: server memory 4 → 10 GB with GC settings, view distance 8, simulation distance 6, on a VPS with a game-server panel.
Result
223 server mods on Minecraft 1.21.1 (NeoForge), measured at 20 TPS (5.3 ms per tick).
Pokémon League
A competitive league mod in Kotlin: gyms, ELO, tournaments.
Context
A full competitive mode on top of Cobblemon, the Pokémon mod for Minecraft.
What I did
- Started the project with another developer and built its economy system.
- The mod covers a 7-stage league with 5 gyms, ranked 1v1 and 2v2 with ELO, double-elimination tournaments for up to 32 players and 96 quests.
SyncCast
Keeps esports casters on the same frame across Twitch and YouTube.
Context
Esports casters commentate the same match from different places — and each platform delivers it with its own delay.
What I did
- Built the product in 8 days (268 commits, December 2025): rooms, a leader caster as reference, and resyncs scheduled on a shared server clock (Ably).
- Twitch drift measured every second; a spread over 500 ms triggers a scheduled realignment.
- An own HLS relay: streamlink + FFmpeg into 1-second segments on Cloudflare R2, segments uploaded before the playlist.
- Teams with Admin / Manager / Caster roles and invites; room chat and live drawing over the video; Row Level Security on every table.
- Redesigned the landing page in 2026 to explain the backend — only with what the code actually does.
Result
Beta, with 52 unit and 14 end-to-end tests.