Measured results
What these systems changed
Not a list of deliverables, but a list of usage that stuck. Every figure below comes from a shipped, measured system, used by teams I do not manage.
of inbound applications decided by AI, with a human kept on the ambiguous cases
hand-decided applications the pipeline agreed with, before go-live
the matching engine judge is capped and parallelised, so cost and latency per brief stay flat as the pool grows
per data extract, self-serve, without going through an engineer
returned to the Product Manager by the human-in-the-loop decision pipeline
more salespeople given access to an expert tool that licences had kept to a few
Selected work
Systems in production
Designed and coded end to end, front and back, shipped and in daily use at Crème de la Crème.
Opening a company back office to natural language
Anyone in the company can ask for something in plain language and have it done, without clicking through five screens, and without ever seeing a row they are not entitled to.
each answering under the rights of the person asking, never those of a service account
Hybrid retrieval, ranking and bounded LLM judgment
A matching engine that had to beat the one already in production, brief by brief, before it was allowed to ship.
a judge layer capped and parallelised, so cost and latency per brief stay flat however large the pool gets
Human-in-the-loop decisioning on inbound applications
The model settles the clear cases. A person keeps the ambiguous ones, and that is exactly what makes the automation acceptable.
of inbound applications settled without a human in the loop
Method
From a business problem to a system people use
Discovery with the people who will use the system, architecture with the trade-offs written down, the code itself, a production path that holds up when the system gets it wrong, and adoption tracked after go-live. Evals, guardrails and cost control are first-class concerns, not afterthoughts.
Discovery
Discovery with the leadership and with the people who will live with the result, turned into a short use-case portfolio: an owner, a success measure, a blocker list. Including the cases worth killing.
Tested against six business teams with different risk tolerances.
Architecture
Models, data, integration, identity, privacy, governance, evaluation, deployment. Every significant call is written down with what it costs.
32 trade-offs recorded before a line of the MCP server was written.
Proof of value
A prototype on real data early, so the decision to go further rests on evidence rather than on a demo that only works on the happy path.
An eval harness that had to beat the production baseline, brief by brief.
Build
I write the code, front and back, working through coding agents, and what ships goes through the client's engineering review. When a piece needs expertise I do not have, I bring in someone who does and write down who owns it once they leave.
Six systems built this way, front and back, on the same engagement.
Production and governance
Evals against whatever is already running, guardrails wired in rather than suggested, authorization and audit designed in. The question of what happens when the system gets it wrong is answered before it ships.
Isolation predictions written down before each role switch, then checked.
Adoption
Usage tracked after go-live, the build documented so another engineer can take it over, and systems retired when a vendor closes the gap.
Rolled out to Sales ahead of general availability.
- 01
Discovery
Discovery with the leadership and with the people who will live with the result, turned into a short use-case portfolio: an owner, a success measure, a blocker list. Including the cases worth killing.
Tested against six business teams with different risk tolerances.
- 02
Architecture
Models, data, integration, identity, privacy, governance, evaluation, deployment. Every significant call is written down with what it costs.
32 trade-offs recorded before a line of the MCP server was written.
- 03
Proof of value
A prototype on real data early, so the decision to go further rests on evidence rather than on a demo that only works on the happy path.
An eval harness that had to beat the production baseline, brief by brief.
- 04
Build
I write the code, front and back, working through coding agents, and what ships goes through the client's engineering review. When a piece needs expertise I do not have, I bring in someone who does and write down who owns it once they leave.
Six systems built this way, front and back, on the same engagement.
- 05
Production and governance
Evals against whatever is already running, guardrails wired in rather than suggested, authorization and audit designed in. The question of what happens when the system gets it wrong is answered before it ships.
Isolation predictions written down before each role switch, then checked.
- 06
Adoption
Usage tracked after go-live, the build documented so another engineer can take it over, and systems retired when a vendor closes the gap.
Rolled out to Sales ahead of general availability.
Capabilities
What I actually build
The parts that decide whether a system can be left running on production data.
Retrieval & ranking
Multi-signal recall (BM25, skill overlap, vector kNN over OpenAI embeddings stored in pgvector, experience indexes) fused by Reciprocal Rank Fusion, then deterministic weighted scoring, then a bounded LLM-as-a-judge layer with constant cost and latency per request. A brief costs the same to answer whether the pool holds a thousand profiles or fifty thousand.
Agents & MCP
MCP servers over real production systems, with OAuth 2.1, tool design and allowlists, sub-agents, packaged skills, and coding agents (Codex, Claude Code) as a normal way of working. Agent Builder and ChatKit where a hosted surface beats a custom one. An agent acts with the rights of the person asking, never those of a shared service account.
Evals & reliability
Golden datasets, eval harnesses that must beat the production baseline before a ship decision, schema-validated structured outputs, and multi-state verdicts scored axis by axis so every decision stays explainable.
Governance & authorization
Three-layer authorization, personas with data-computed scopes, field-by-field payload projection, deterministic GDPR guards outside the model path, and queryable audit trails with bounded retention. What a departing employee can still see is answered with a date rather than a guess.
Stack
TypeScript and Python, React and Node, PostgreSQL with pgvector, Vercel, PostHog. Coding agents are how I build: the architecture is mine, what they produce is corrected rather than accepted, and what ships is reviewed by the client's engineers. The Responses API, structured outputs and embeddings run in production here; several frontier providers run side by side, picked per task rather than by default.
Adoption
Discovery with non-technical owners, documentation and handover, usage tracked after go-live, and systems retired when the vendor already in place closes the gap.
Modularcomponents,oratoms,recombinedintolargersystems.ItiswhyInamedmycompanyAtomly AI.
Trusted by
Teams I have worked with
Twelve years of client-facing work: technical discovery with executives, GenAI proofs of concept taken into production pipelines, and 1,000+ professionals trained in applied GenAI.
What I do
Applied AI, from the business problem to production
I work as the technical owner of an AI portfolio: discovery with the people who will use the system, architecture with the trade-offs written down, the code itself, production, and then adoption. Three ways to work together.
Testimonials
What clients say
From the enablement side of the work, the part that decides whether a system survives its first month.
“Pierre is an exceptional trainer who knows how to make generative AI exciting and accessible to everyone. His depth of expertise, combined with a rare ability to adapt to participants' industries and professions, makes each session a tailored experience. His training sessions are dynamic, inspiring, and always perfectly aligned with clients' concrete challenges. With attentive follow-up and highly precise feedback, Pierre combines professional excellence with the joy of learning!”
Christophe Conceicao
Head of Education & Operational Excellence, Maria Schools
“Beyond being one of the best AI experts today, Pierre is a natural teacher, making his training sessions both useful and interesting! He conveys his passion for AI with energy and conviction, and consistently tries to relate examples to cases from the participants' industry.”
Guillaume Coulomb
Co-founder, On Train
Engagements
Recent engagements
Beyond the production systems: assessments, workshops and training programmes delivered to technical and business teams.
Crème de la Crème
Technical owner of the AI portfolio: discovery with the Sales, Product and Engineering teams, architecture, code and production. MCP servers, a hybrid matching engine, human-in-the-loop decisioning, governed data access.
Direct client
BHG Conseil
Executive coaching on generative AI use cases, focusing on optimising various consulting-related tasks. In-depth exploration of AI Agent creation to automate and enhance consulting processes.
Direct client
Cube School
Created a comprehensive e-learning program with over 10 hours of video content. Developed a structured curriculum providing a complete introduction to generative AI for students.
Direct client