From LLM integration to voice AI at scale — I deploy production-ready AI systems in the field, not proofs of concept.
Paris, France — remote or on site1 000 €/dayAvailable for freelance missions
Status
Available for projects
5M+
Daily requests
served in production
Deployed for
Not demos, not POCs — production AI systems handling real traffic, real users, real failure modes.
Serving 5M+ manga reads daily.
Built from scratch · High-traffic platform · 5M+ daily requests
Built and deployed LLM-powered voice agents at scale.
Built and maintained in production over 7 years · Voice AI systems · Real-world deployments
Translate text inside images — instantly.
Built from scratch · Live in 2025 · Real-time image translation
Automating Dofus forgemagie with real-time updates and reverse engineering.
Built from scratch · Reverse engineered · Auto-updating with game changes
A swipe-based job discovery app designed to make matching faster and more intuitive.
Built from scratch · Live in 2026 · First users onboarding
AI chat widget that captures and qualifies leads through natural conversation.
Built from scratch · Live · 9 languages · Multi-source · Drop-in embed
The global super wallet for Web3.
Built from scratch · Live in 2026 · Web3 · Social + Finance convergence
7+ years deploying production AI — voice agents at talkr.ai, computer vision pipelines, LLM APIs — across systems that handle millions of requests daily.
Forward deployment means being the engineer who actually makes it work in the field. Not a POC specialist — the one on the ground when it hits production.
I know what breaks under real traffic, how latency compounds at scale, and how to architect AI systems that stay reliable when the demo ends.
How I operate
LLM · Voice · Vision
I wire LLMs, voice models, and vision pipelines into real products — not sandboxed experiments. From prompt engineering to production API design.
OpenAI · Anthropic · Whisper · ElevenLabs · MCP
Field notes from production systems — voice agents, LLM pipelines, scaling decisions and the ones I got wrong. Written for whoever has to make the call, not only for engineers.
A client needed software that follows someone toward an objective for months, not a chatbot that answers for ninety seconds. Here is what that changed in the architecture — and the four decisions I would defend again.
The concrete optimizations that took opus from janky to silky-smooth on sub-$200 Android devices — FlatList tuning, JS thread management, Hermes, and the things that actually move the needle.
How I built manju's real-time image text extraction and translation system — the technical decisions behind the OCR pipeline, model selection, latency optimization, and handling text in complex image layouts.
Real lessons from 7 years building and scaling a voice AI platform in production — context windows, latency, edge cases, and why most voice demos fail in the real world.
Most AI projects stall between demo and production. I close that gap — because I've been on the ground when it matters.
Before any integration, I map the constraints: latency budget, data sensitivity, existing stack, failure tolerance. The right architecture starts with the right questions about the environment.
A demo is a controlled lie. Production has unpredictable load, edge cases, cost ceilings. I design AI systems that degrade gracefully and recover fast — not ones that impress in a slide.
The fastest path to a working system is shipping to real users early. I deploy v1 in days, then instrument everything — latency, fallback rates, token usage — and iterate on what actually breaks.
Forward deployment doesn't end at go-live. I stay until the system is observable, the team is autonomous, and the failure modes are documented. I don't disappear after the handoff.
Most people build features.
Few build products people come back to.
I focus on shipping fast, iterating from real usage, and turning ideas into products that reach scale.
A production AI integration, a voice agent, a system that needs to handle real scale — I work on deployment missions where the stakes are real. No forms: just a direct email.
