Benoit Barthelet

benoit.barthelet@gmail.com • +33 698 609 237 • 48 years old

13, rue d’Alexandrie - 75002 Paris, France

AI & LLM engineer who ships — builds production AI workflows end-to-end: from self-hosted LLM inference (vLLM, Ollama, llama.cpp) and RAG pipelines to model routing, agentic workflows, and prompt engineering. Author of deepseek-acp-adapter, bringing open-weight reasoning models into the editor. Uses AI as a force-multiplier daily and designs tooling so others can too.
Finance-native engineer — 15+ years in capital markets (portfolio management, derivatives pricing, risk) combined with hands-on platform engineering. Understands both the operational reality of a trading floor and how to ship software that holds up under scrutiny in highly regulated environments.
Builds tooling for how people actually work — former maintainer of uvicorn, the ASGI server underpinning most Python backends; contributor to Litestar. Ships small, focused tools that solve real friction.
Thinks in systems, not just code — from Terraform IAC to CI/CD pipelines, from API design to AI-powered workflows, comfortable owning the full stack. OSCP-certified, so security isn’t an afterthought.
Operates at every altitude — equally at home defining architecture with stakeholders, writing async Python at 2am, or explaining tradeoffs to non-technical decision-makers in terms they care about.


AI/LLM Engineering & MLOps	Cloud Infrastructure (AWS, GCP)	Capital Market Products & Processes
API design & Backend Architecture	Data Pipelines & Migrations	Team Enablement & Coaching
Python / asyncio (expert)	PostgreSQL (advanced)	Open-Source Development & Maintainership

Selected projects

AI & LLM

deepseek-acp-adapter — github.com/euri10/deepseek-acp-adapter Open-source protocol adapter bridging DeepSeek reasoning models (R1, V3) to any ACP-capable editor (Neovim, VS Code, Zed). Built to break free from closed, vendor-tied coding assistants. Implements the Agent Communication Protocol for real-time streaming, tool-use, MCP servers and multi-turn conversations.

RAG-powered Financial Intelligence Assistant (SFJ Technologies) End-to-end retrieval-augmented generation system for natural-language querying of proprietary financial datasets. Architected the full pipeline: embedding generation, chunking strategies, vector search (pgvector), hybrid retrieval logic, and LLM orchestration via LiteLLM with model fallback and cost-aware routing across OpenAI, Anthropic, Perplexity.

Local LLM Infrastructure Designed and operate a home-lab inference stack running open-weight models (Llama 3, Qwen, DeepSeek, Mistral) on consumer GPU hardware using vLLM, Ollama, and llama.cpp. Uses this as a proving ground for model evaluation, prompt engineering, and agentic workflow patterns before bringing them into production contexts.

Infrastructure & Open Source

uvicorn — uvicorn.org Former maintainer of the lightning-fast ASGI server built on uvloop and httptools. Powers a significant fraction of Python web services in production today.

litestar — github.com/litestar-org/litestar Contributor to the high-performance ASGI framework. Deep familiarity with the modern Python async stack (msgspec, uvloop, httptools).

Experience

SFJ Technologies LLC − Chief Architect and Head of Engineering − Sep 2018 ‣ Present

Fintech generating signals to measure and detect bankruptcy and fraud risks. Scaled the technology environment by deploying enterprise-strength operations and a microservice architecture, all within a regulated data-governance framework.

AI & LLM Engineering

Architected and shipped an internal RAG-powered assistant for natural-language querying of proprietary financial data — embedding generation, vector search (pgvector), hybrid retrieval, and multi-model LLM orchestration via LiteLLM with cost-aware routing across OpenAI, Anthropic, and self-hosted models.
Established LLM evaluation pipelines: prompt versioning, A/B testing of model responses, retrieval quality metrics, and hallucination guardrails for finance-domain outputs.

Platform Engineering

Architected the back-end platform: FastAPI / PostgreSQL / arq, with CI/CD on GitLab and IAC via Terraform.
Designed API consumer surfaces including Excel User-Defined Functions for users who live in spreadsheets.
Built data pipelines converting raw accounting datasets into scored signals through ML algorithms — ingestion, transformation, delivery.
Single point of technical ownership: infrastructure, security posture, data residency, and delivery.

Sullivan Cloud − Chief Solution Architect − Sep 2017 ‣ Present

Leading advisory and system integrator for digital transformation of legacy workflows. Designed and developed cloud migration tooling on GCP.

Built automation tools for GSuite migration across manufacturing and professional-services clients.
Reviewed and hardened cloud infrastructure for a broad range of industries.

Generali Investments Europe − Portfolio Manager − Dec 2010 ‣ Sep 2017

Actively managed a multibillion portfolio of derivative-based assets and drove the technology transformation of the front-office toolchain.

Created insurance-specific investment strategies on European fixed-income derivatives.
Developed in-house pricing and risk-modelling software, replacing black-box vendor systems with auditable, self-built tools.
Automated end-to-end front-office processes — from trade capture to risk reporting — giving the desk a speed advantage in opportunistic markets.

Lazard Freres Gestion − Portfolio Manager − Jun 2005 ‣ Dec 2010

Created the structured-product and Fund-of-Hedge-Funds desk for ultra-high-net-worth clients.

Defined product strategy, pricing, and marketing for a new business line.
Acted as principal architect for the technology platform supporting the full product lifecycle.

SGAM Alternative Investments − Fund derivatives trader − Mar 2001 ‣ Jun 2005

Pioneered a new market in hedge-fund derivatives, leading to the first billion-euro capital-guaranteed mutual fund in Europe.

Managed a €1B book of highly illiquid fund derivatives.
Built pricing models and an integration platform that automated workflow across sales, middle-office, back-office, and IT.

Education & certifications

Ecole Supérieure de Commerce de Paris — MSc Finance, Derivatives Trading Thesis: pricing of PCS options (Property-Claim Services) and impact on reinsurance natural-disaster markets

OSCP — Offensive Security Certified Professional (OS-101-035377)

GitLab Certified Associate

Associate Google Cloud Engineer

Skills


LLM & AI	RAG pipeline architecture, embedding strategies & vector search (pgvector), LLM orchestration (LiteLLM), multi-model routing & fallback, prompt engineering, agentic workflows, LLM evaluation & guardrails, ACP (Agent Communication Protocol)
AI tooling	Neovim + CodeCompanion, custom DeepSeek ACP adapter, self-hosted open-weight models (vLLM, Ollama, llama.cpp) on local GPU, model evaluation harnesses
Languages	Python (expert, deep asyncio), JavaScript / React, C# (Excel-DNA), Rust (learning)
Infrastructure	AWS, GCP, Terraform, Docker/Podman, GitLab, Debian daily driver
Databases	PostgreSQL (advanced), async migrations, pgvector
Security	OSCP-certified; threat modelling for web applications and APIs; AI safety & prompt injection awareness
Languages (spoken)	French (native), English (fluent), Chinese (beginner)

Personal

Home Lab — Self-hosted GitLab with CI/CD pipelines for all personal services; local LLM inference farm (Ollama, llama.cpp, vLLM) running open-weight models on consumer GPU — used for continuous experimentation with model routing, agentic patterns, and RAG architectures.
Ultratrail
Gardening

Debugging is twice as hard as writing the code in the first place.
So if you’re as clever as you can be when you write it,
how will you ever debug it? – Brian Kernighan

Inspired by pandoc_resume