Pierre Aumjaud · Freelance AI engineer
From vision pipelines and AI agents to the web application your team uses every day — designed, built, and shipped by one engineer you can talk to.
I’m Pierre Aumjaud, the engineer behind Seed AI. I moved into freelance AI engineering from research: a PhD, peer-reviewed publications, and years of building and evaluating machine-learning systems before putting them in front of users.
Today I build AI tools that hold up in production — like FootCheck, a match-sheet analysis application for Helicon IA — and the multilingual, content-driven websites that businesses maintain themselves.
My full technical background — CV, publications, and engineering blog — is at pierreaumjaud.com.
Repetitive document work, automated: vision models that read photographed or scanned documents, extract structured data, and flag what needs a human decision. For Helicon IA, FootCheck reads football match sheets and identifies suspended players — replacing a manual check for every round.
See related work →Assistants that do real work under your control: RAG systems grounded in your own documents with cited sources, email and calendar agents with human-in-the-loop approval before every action, on local or hosted models — with graceful fallbacks when the model is unavailable.
See related work →From a multilingual site your team edits without a developer to a full-stack application with authentication and an API. This site is itself a deliverable: a custom Hugo build, bilingual, accessible, statically hosted — view the source if you like.
See related work →
Client project · Built for Helicon IA
Match sheets checked against suspension lists — automatically.
Football officials must check every match sheet against the week’s suspension list — a slow, error-prone manual task repeated across a whole round of matches. For Helicon IA, I built FootCheck: a full-stack web application where match sheets are uploaded as photos or scans, a vision-model pipeline extracts the players’ names, and suspended players are flagged automatically. A round is now processed about five times faster than by hand, with 97% accuracy in identifying suspended players.
Personal project
Ask a textbook questions; get answers with cited sources.
Finding the right passage in a dense reference document takes longer than understanding it. This assistant ingests PDFs — with OCR and table extraction — into a persistent ChromaDB vector index and answers questions grounded in the source, citing the exact passages it used. Retrieval uses maximal marginal relevance to avoid redundant context, and providers are swappable through configuration (Ollama, OpenAI-compatible APIs, vLLM). The same architecture applies wherever a team needs reliable answers from its own documents: manuals, procedures, contracts.
Personal project
An inbox agent that classifies, drafts, and schedules — and asks before it acts.
Inbox triage follows predictable patterns — but nobody wants an AI sending mail unsupervised. This agent classifies incoming email with confidence scores, drafts replies with tone control, and proposes meeting slots from calendar availability; every outgoing action requires explicit approval, risky items are flagged for review, and each decision is stored for audit. When the model is unavailable, it degrades to keyword rules and templates instead of failing. The pattern — automation with human sign-off — fits any workflow where mistakes are expensive.
Personal project
Fast, multilingual sites their owners update themselves.
Four personal projects — not client engagements — showing one capability: content-driven sites that stay fast and that their owner maintains without a developer. One is designed and built from scratch in Hugo, bilingual with a language switcher and anchor navigation; the other three are Jekyll builds, customised from existing themes into a topic-hub ML blog, an editorial blog with article series, and an academic portfolio.
Bilingual EN/FR Hugo single-pager: anchor nav, language switcher, pricing tiers, integrated blog. Designed and built from scratch.
Machine-learning blog organised into seven topic hubs. Jekyll, built and customised.
Editorial blog with category navigation and multi-part series. Jekyll, built and customised.
Personal and academic site: CV, portfolio, publications, blog. Jekyll, built and customised.