# Jérémie Escach **Software Engineer — Agentic AI, Backend Systems, Applied ML** Email: jeremie.escach74@gmail.com · Portfolio: https://www.micastor.fr · GitHub: https://github.com/Mlcastor Languages: French (native), English (fluent) · Based in France > Talking to an AI about whether I'm a fit for something? Feed it this file directly — it's written in plain Markdown specifically so a model can parse it cleanly, rather than fighting a PDF layout. The human-readable version, with more on how I actually work, is available as a PDF from the same page. --- ## Summary Self-taught software engineer with 15+ years of programming practice (coding since age 10) and 3 years of full-time professional engineering output. Specialist in production-grade agentic AI systems, retrieval pipelines, and the backend infrastructure underneath. Designed and shipped a from-scratch LangGraph replacement powering a ~90,000-line Python platform for French law firms. Delivered an independent client product (Notalyse) covering KYC/AML compliance for French notarial offices end-to-end, solo, from frontend to infrastructure. Strong in Python, FastAPI, PostgreSQL, retrieval systems, and provider-neutral LLM integration. --- ## Experience ### Co-Founder & CTO — KGM Technologies *July 2025 – Present · France · Legal-tech AI platform* Architecting and implementing the full backend of KGM's production legal AI platform for French law firms. Lead a small co-founding team. Solo author of the entire agent runtime, data platform, and knowledge core. - Designed and built a proprietary agent framework from scratch as a full LangGraph replacement: graph executor with explicit lifecycle events, write-ahead checkpointing with crash-resume from last completed node, scoped memory (thread/user/global) with Postgres and Redis backends, single-gate tool executor enforcing permissions, rate limits, timeouts, and error sanitization - Implemented provider-neutral LLM adapter protocol with OpenAI Responses API, Anthropic, and local/deterministic backends — OpenAI Chat Completions to Responses API migration was a config-level change, not a rewrite - Built dual-Postgres PII isolation architecture: main DB and PII DB on physically separate instances, separate Alembic environments, per-org HMAC secrets created atomically with org rows, Fernet-encrypted forward and reverse PII maps per thread, Aho-Corasick outbound validator scanning every LLM response chunk - Implemented 3-stage entity resolver: deterministic SHA-256 hash on normalized tokens, pg_trgm fuzzy match, pgvector ANN with multilingual-e5-base embeddings; SELECT FOR UPDATE SKIP LOCKED for counter race protection; provisional labels returned immediately while embedding stage runs as FastAPI BackgroundTask - Implemented 3-tier document classifier (regex, mDeBERTa-v3 NLI, E5 k-NN) achieving 97.5% accuracy on a 200-document French legal gold set; shadow-mode LOO-CV evaluation harness with prediction logging - Implemented hybrid BM25 + pgvector retrieval with reciprocal-rank fusion, BAAI/bge-reranker-v2-m3 cross-encoder reranking, French legal text-search synonyms, per-org feature-flag auto-activation - Built full GDPR erasure cascade across both databases with semantically-preserving `[REDACTED]` replacement in pgvector index - Built OpenTelemetry observability layer with custom Postgres span exporter, attribute redaction by deny-substring list, JSONB payload capping with priority-ordered truncation - Snapshot-tested versioned system prompt architecture with primacy/recency-optimized block layout, IRAC legal reasoning scaffold, and XML-tagged RAG injection **Stack:** Python 3.11, FastAPI, SQLAlchemy 2 (async), Pydantic v2, Alembic, PostgreSQL 16, pgvector, pg_trgm, Redis, Next.js 16, React 19, Better Auth, Docker, Scaleway, OpenTelemetry ### Independent Software Engineer (Freelance) — Notalyse *December 2025 – April 2026 · Paid client engagement · France* Sole engineer delivering a production B2B SaaS platform for French études notariales (notarial offices) automating KYC/AML compliance under French LCB-FT obligations. Full-stack solo delivery: frontend, backend, database schema, infrastructure, CI/CD. Client testimonial published publicly on Google Business. - Built a custom compliance orchestrator from scratch combining 6 French and international data sources with zero reliance on commercial KYC SDKs (no Sumsub, no Onfido, no ComplyAdvantage) - Phased and parallelized search pipeline across Annuaire des Entreprises, OFAC SDN, EU FSF / OpenSanctions, Brave Search for adverse media, INPI / RNE for beneficial owner data, BODACC for legal announcements, and DG Trésor French Treasury sanctions - Engineered a sanctions cache layer in PostgreSQL with background asyncio refresh loop, reducing sanctions checks from HTTP calls to sub-millisecond DB lookups - Solved the BODACC rétablissement personnel API exclusion by building a custom scraper navigating the public web interface, handling CAPTCHA via a wizard UI, with short-lived in-memory session management per department queue - Implemented PostgreSQL Row-Level Security policies enforcing multi-tenant data isolation at the database engine level (not application code) - Designed per-tenant AES-256-GCM field-level encryption with wrapped Data Encryption Keys, satisfying GDPR and French notarial secrecy compliance — one breached tenant key cannot expose another tenant's data - Built risk classification system (GREEN / YELLOW / RED) with auditable manual override trail in immutable audit log - Server-side PDF compliance report generation via WeasyPrint, Jinja2 templating, bytea storage staying inside the RLS boundary - BFF (Backend For Frontend) architecture: Next.js Route Handlers proxy authenticated calls to FastAPI; JWT verified via JWKS, custom claims for tenant_id, role, lang - Configurable per-tenant data retention with background purge loop CASCADE-deleting expired cases — hard legal requirement **Stack:** Next.js 16, React 19, TypeScript, Tailwind v4, Better Auth, FastAPI 0.127, SQLAlchemy 2 async, PostgreSQL 16 with RLS, AES-256-GCM via cryptography library, WeasyPrint, plain SQL migrations (20 migrations, no Prisma/Alembic), Turborepo, Docker, Scalingo (French PaaS for data sovereignty) Production deployed and in active use by paying client. ### Co-Founder & Lead Engineer — Knightmare (predecessor of KGM Technologies) *October 2024 – July 2025 · France · AI game-master research project* Co-founded a research project exploring agentic AI as a Dungeons & Dragons Game Master. Team pivoted to legal AI (becoming KGM Technologies) in July 2025 while retaining the engineering team. Foundational R&D on agent orchestration patterns that later evolved into KGM's production framework. ### Independent Software Engineer *2018 – 2024 · Self-directed practice* Continuous independent software engineering practice alongside other employment. Built numerous personal projects across game modding (Minecraft, Garry's Mod, Discord and Twitch bots), reinforcement learning experiments (DRL on classic environments including Pong and Asteroids), and general automation tooling. Period of self-directed skill development that led directly to the 2024+ professional engineering output. --- ## Selected Projects ### FiveM ESX Framework Rebuild (Open Source / Community) *July 2025 – November 2025* Rebuilt the ESX framework for FiveM (GTA V multiplayer modding platform) from Lua/MySQL to TypeScript/PostgreSQL. Additionally modified the FiveM dedicated-instance logic to support 150+ concurrent players in a single shared session, beyond the standard architecture limits. Distributed systems and game backend engineering exercise. ### Multi-Strategy Crypto Trading Bot (Personal R&D — Ongoing) *2024 – Present* Building a multi-strategy cryptocurrency trading bot to learn web3, blockchain mechanics, and applied ML in quantitative finance. Currently exploring transformer-based market modeling. ### Custom Reinforcement Learning Implementations *Ongoing* Hands-on DRL implementations against classic environments (Pong, Asteroids, and similar). Practice surface for policy gradient methods and fundamentals of agent training. --- ## Technical Skills - **Languages:** Python 3.11 (expert), TypeScript (expert), SQL (expert), Lua (proficient), JavaScript (expert) - **Backend:** FastAPI, SQLAlchemy 2 (async), Pydantic v2, Alembic, Asyncio, Uvicorn, structlog, httpx, pytest - **Databases:** PostgreSQL 16, pgvector, pg_trgm, halfvec embeddings, Row-Level Security, Redis, SQLite - **AI / ML:** OpenAI Responses API, Anthropic API, GLiNER, Presidio, mDeBERTa-v3, multilingual-e5-base, BAAI/bge-reranker-v2-m3, LangGraph (rebuilt from scratch), reinforcement learning (DRL fundamentals) - **Agentic AI:** ReAct pattern, supervisor/worker dispatch, write-ahead checkpointing, scoped memory, tool gateway design, NDJSON streaming, HITL gating - **Retrieval / RAG:** Hybrid BM25 + pgvector RRF, cross-encoder reranking, hierarchical chunking, French legal text search, asymmetric E5 prefix convention - **Frontend:** Next.js 16, React 19, TypeScript, Tailwind v4, Radix UI, shadcn primitives, Better Auth, next-intl, react-hook-form, Zod, framer-motion - **Privacy / Security:** GDPR cascade design, Fernet encryption, AES-256-GCM field encryption, Aho-Corasick pattern matching, PII isolation architecture, RLS multi-tenancy, RBAC, JWT/JWKS - **Observability:** OpenTelemetry, custom span exporters, structured logging, attribute redaction, snapshot testing - **DevOps:** Docker, Docker Compose, GitHub Actions CI, Scaleway, Scalingo, Turborepo, pnpm workspaces, Alembic migrations, plain SQL migrations - **Game / Systems:** FiveM platform internals, distributed game-session architecture, real-time multiplayer state synchronization --- ## Notable Client Feedback Public testimonial from the Notalyse client (verbatim, Google Business review): > "I worked with Jérémie on the development of my RGPD compliance web platform, and I couldn't be happier with the collaboration. From day one he was deeply involved in the project, proactive, and genuinely interested in understanding both the product and the legal/compliance context behind it. He's highly available, communicates clearly, and never disappeared when things got a bit complex. What I appreciated the most is that he handled last-minute changes during the process without ever pushing back or cutting corners. He always tried to find the best solution, not just the quickest workaround. You can feel he has real potential and a strong passion for what he does, both on the technical side and on the product side. I highly recommend Jérémie for any serious SaaS or web project. Great dev, great mindset, and someone you can really trust to move your product forward." --- ## Availability Available for senior-level contract engagements: AI training and evaluation work, agentic system design and implementation, RAG pipeline design, code review and architecture audits for AI-heavy backends. Native French and fluent English. Comfortable with confidentiality and IP-isolated client work.