10/08/2026 updated
Premium member
100 % availableAI Engineer
Villiers sur Marne, France
Only remote
AI AgentsRAGMLOpsLangChainArtificial IntelligencePython (Programming Language)Software ArchitectureNext.jsTypeScriptReact.jsMulti-Agent SystemsFastAPI
Freelance AI Engineer. I build production agent systems: multi agent orchestration, MCP (Model Context Protocol) servers, industrialised RAG, tool calling, and the evaluation and observability layer that makes those systems measurable. Python and FastAPI on the back end, TypeScript and React on the front.
I also have a full stack product background: three products designed and shipped to production on my own, back end through deployment. That combination, engineering plus the ability to move from idea to working software, is what I bring.
What I do
AI agents in production: multi tool agents with state persisted in PostgreSQL, error handling and resumption, guardrails verified by a test on every build.
Architecture: LLM pipelines built as a graph driven by an explicit state machine; multi model routing with two stage provider fallback, each link gated by its own circuit breaker; exponential backoff with jitter, idempotency keys, schema validation at every stage, streaming.
RAG, production grade: multi source ingestion, semantic chunking, embeddings, BM25 plus dense plus hybrid retrieval, neural reranking, pgvector and Qdrant.
MLOps and reliability: automated evaluation sets, hermetic offline test suites, LLM tracing and observability, cost per stage, p50 and p95 latency.
Python and FastAPI first: SQLAlchemy, Celery task queues, PostgreSQL 16 with pgvector, Redis 7, Docker and docker compose, CI/CD on GitHub Actions, containerised deployment in EU regions.
LangGraph, TypeScript and React: LangChain and LangGraph; Next.js 15 App Router with React 19, Tailwind 4, Supabase (RLS, Auth, Storage); Rust and Tauri 2 on desktop.
MCP servers: servers and clients exposing tools to MCP aware clients, personal token authentication, server side input validation.
Products I shipped on my own
A scoping and quoting SaaS running a four stage LLM pipeline in production, with PDF and OCR ingestion, DOCX and PDF exports, advanced eIDAS signature and Stripe billing. A signed and Apple notarised desktop dictation application for macOS and Windows, distributed with minisign verified updates. A Next.js platform for niche research and video production support.
Two open source agentic building blocks: a shared memory for AI agents exposed over MCP, and a patent claim scope drift auditor with verifiable provenance on every assertion, covered by 143 hermetic tests.
How I prove it
Reproducible measurements produced on my own products, on synthetic data: a bias audit over 50 names across 5 origins, three runs, FAIL verdict, 39.8 percent maximum spread between groups, framed against EU AI Act article 10; 88 percent of pipeline cost concentrated in 2 of 4 stages, which decided the routing.
Engineering degree from CentraleSupelec, 2024.
Code: github.com/AmirK-S - Site: amirks.eu
How I work
Fully remote, zero days on site. Available immediately. Working languages: French (native) and English (C1 Linguaskill, Cambridge).
Experience
Professional experience0 - 5 years
Freelancer since2024
Languages
EnglishFluentFrenchNative speaker
Project history
Own SaaS product, designed, built and operated single-handed: getscope.dev turns a raw brief (RFP PDF, scanned tender, meeting transcript) into a costed scoping document. Four-stage LLM pipeline driven by a state machine; multi-model routing with two-tier fallback, per-provider circuit breaker, exponential backoff, idempotency, Zod validation, streaming. PDF, OCR and audio ingestion; DOCX and PDF export; self-hosted eIDAS electronic signature; Stripe billing. MCP (Model Context Protocol) server exposing the product to Claude Code and Cursor. Six CI jobs and 30 end-to-end scenarios. In production.
Own product, designed, built and released single-handed: TTP (Talk To Paste), a desktop dictation app. Native macOS and Windows app (Tauri 2, Rust and React), version 3.1.6 shipped, paid licences. Transcription in under two seconds, LLM clean-up pass, personal dictionary. Apple-signed and notarised builds, signed Windows installers, minisign-verified updates, API keys stored in the OS keychain, published threat model.
Own product, designed and developed single-handed: Next.js platform (VideoForge / NicheRadar) for niche research and video production assistance. 12-module technical masterclass (140+ pages) produced and published.