Loïc Ledreck

Software engineer in Paris. By day, software engineer at Zefir. The rest of the time, I run Tablia — a company whose workforce is AI agents.

My conviction is that intelligence is no longer the bottleneck, judgment is. I build the context, verification loops and escalation paths an AI agent needs to exercise judgment inside a real organization.

Tablia — founder Zefir — product engineer Swan — 3.5 yrs, banking-as-a-service
Now

Tablia

A reservation platform for restaurants — back office, booking widget, guest notifications, Google integration — at a fraction of the incumbents' price. In production since June 2026; the first restaurants are live and taking real bookings. Every feature is specified, built, reviewed, tested and deployed by a pipeline of Claude agents that I orchestrate.

tablia-resa.fr Book a table — see it live Reservation widget · back office · in production

The real widget, live from production — bookings land on the demo restaurant.

The harness behind it

Eight specialized agents — PM, designer, staff engineer, QA, builder, three independent review lenses, vision QA — working over a documentation layer that acts as the company's shared memory. 25 pipeline runs in nine weeks, 17 of them fully autonomous, launched at night. Median 4h48 of agent-active time per feature.

blog.loicledreck.com Judgment is a harness property The write-up — 7 min read
AI in production at Zefir

ZIA, a lead-qualification chat agent at the top of Zefir's growth funnel: it onboards property sellers through conversation, delivers an instant price estimate and answers with dynamically generated UI widgets — 8+ tools, in production.

A document-validation pipeline for the paperwork behind every real-estate mandate: OCR plus LLM checks against business criteria, with prompt suites evaluated in Braintrust.

Claude skills that automate a codebase-wide ORM migration (TypeORM → Drizzle).

Enterprise engineering at Swan

Verification of Payee for Swan's payment systems, ahead of the EU regulatory deadline — designed, implemented, then coordinated the API rollout with Swan's 100+ enterprise partners.

Migrated the webhook service — 10,000+ events per hour, one of Swan's most critical — from DynamoDB to PostgreSQL with zero downtime, moving to a queue-per-subscription model.

Built the CI/CD pipeline for a 30-developer organization and drove its adoption: an upstream open-source patch, GitLab templates, docs and team-by-team onboarding.