Freelancer profile translated to English.

Description

PRODUCTION-READINESS AUDITS FOR GITHUB PROJECTS.

Anthropic CVP-verified security researcher · Pre-merge OSS audits · Orionix

I audit OSS code in pre-merge for fast-shipping tech startups. Approved in Anthropic's Cyber Verification Program (June 2026), I operate under the Orionix brand.

What I deliver: 1 note per PR with findings file:line + proof-by-contrast + suggested fix. Zero hallucinations, zero extrapolation. Each claim traces back to a verifiable line of code.

Public track record (verifiable on GitHub):

★ Anthropic CVP approved researcher (June 2026)

- hatchet-dev/hatchet #4131 → PR #4137 merged with my review

- BigBodyCobain/Shadowbroker #375: 3 findings, maintainer fixes in 34 minutes

- Vercel ai-sdk: 3 HackerOne reports, triage feedback: "excellent documentation quality"

- BerriAI/litellm: reporter credit GHSA-wr37-fm8c-4vvq

- Sentry: disclosure CWE-682 under triage (June 2026)

Industry field of expertise

Languages

French
Native or bilingual
English
Fluent

Workplace preferences

Remote only

Primarily works remotely

Orionix
Founder · Production-Readiness Auditor
June 2026 - Today (2 months)
Orionix audits GitHub projects before they break in production. Not "it compiles" — race conditions under load, timeouts that block an entire stack, secrets leaking into logs, missing audit trails that make any postmortem impossible.

Recent public audit — BigBodyCobain/Shadowbroker (9k stars on GitHub). Issue #375, three production bugs identified in 45 minutes — the maintainer shipped the fixes in 34 minutes (commit 9a0a9a1, github.com/BigBodyCobain/Shadowbroker/issues/375).

METHOD
• Deep reading of the backend + state-of-the-art on the stack
• Orionix Pattern Library (60+ production-readiness patterns accumulated audit after audit)
• Deterministic verifier: each finding is re-verified against the AST before publication
• Rationing: 3 public bugs + 2-3 in reserve for paid follow-up

DELIVERY
• Detailed report (file:line + concrete fix) — for standard audits
• Direct refactoring (clone + fix + tests + clean handover) — for premium

PROPRIETARY TOOLS
• Beaume — RAG engine, 672K docs, 90% recall@10, deterministic verifier with 0 hallucination on numerical thresholds. Three-brain architecture, 100% local.
• Jars — native iOS app (5500 lines of code, 102 tests, StoreKit 2, Family Controls) built with the same rigorous method.

Stack: Python, TypeScript, Go, Rust, Swift, orchestrated AI agents.
Python Code Audit Application Security Code Review
Beaume
Founder & Architect
February 2026 - Today (6 months)
Beaume is a 100% local legal RAG engine for French lawyers. First version on labor law (economic dismissal). Three-brain architecture (creative LLM + deterministic expert system + distributed multi-agents), 672K Legifrance documents indexed, deterministic verifier achieving 0 hallucination on numerical thresholds.

Differentiator: we make a deterministic expert system reason; the LLM is only used for natural language translation. The opposite of the Harvey / Doctrine model.

FIGURES
• 672K compacted Legifrance docs (~60 MB free tier)
• 90% recall@10 on BGE-M3
• 28/28 document_analyzer tests + 370/370 non-LLM tests passed
• Runs on Mac Air M2 16GB (free tier)

ROLE
• Product and architecture vision
• AI agent orchestration for development and operations
• OSS strategy (BSL 1.1 level 2) and go-to-market for Lyon lawyers
• YC communication

STACK: Python, RAG (BGE-M3 + FAISS + PageRank), Swift (macOS HUD), Ollama (Gemma 4), orchestrated Claude agents.
https://github.com/mathieuballotma-sketch/lucie