Agentic and AI systems
AI systems that reach production
RAG, assistants and agentic workflows where the interesting part is not the model. Auditable business logic, a deterministic core, and a deployment that survives contact with real users.
Where the LLM sits
Business rules first, a deterministic core second, the language model in narrow slots: reading intent, writing the rationale. Every decision can be traced to a rule or a score, which is what makes it operable.
What this usually looks like
- The demo was excellent. Nine months later nothing is in production.
- The assistant is fluent and occasionally, confidently, wrong, with no way to audit why.
- Every answer requires the LLM, so cost and latency scale with traffic and nobody can explain either.
- Nobody can say what the system will do next Tuesday, because nothing is deterministic.
What you get
A deterministic core
The LLM handles language. Filtering, ranking, business rules and validation are code you can read, test and reason about.
Auditable decisions
Every recommendation carries its rationale and is reproducible. When someone asks why, there is an answer that is not "the model said so".
Control without a deploy
Business owners change what the system promotes through configuration, not an engineering ticket.
Deployed where you need it
Your cloud or on-premises. Monitoring, evaluation harness, and the documentation to operate it.
Selected work
Generative AI
AI-guided recommendation engine for premium floral e-commerce
A production-oriented recommendation system that guides customers through emotionally loaded floral purchases — using a deterministic state machine with LLM components constrained to intent parsing and rationale generation only.
Guided purchase flow in production: auditable recommendations, business rules editable without a deploy
Risk & Operations
Supplier risk intelligence framework
Network-based supplier risk modeling that identifies systemic vulnerabilities and "too-central-to-fail" suppliers before disruptions occur.
Too-central-to-fail suppliers identified before disruption, prioritised by systemic impact
Depth
Jul 28, 2026
One training monorepo, or a repo per model?
Seven ML products, one retail team, and the question of whether to consolidate training into a single Vertex-native monorepo. The answer is conditional, and the condition is the whole point.
May 18, 2026
The Hidden Cost of Precision in Operational AI
Enterprise AI systems often work at the first level of granularity, then become fragile when the business asks for more precision. A field lesson from pharmaceutical supply optimization on why incremental architectures matter.
May 13, 2026
Real Latency vs Perceived Latency in GenAI Systems
Raw latency and perceived latency are different engineering problems. Production GenAI systems feel fast when they expose progress early, overlap backend work, and avoid silent waiting.
May 4, 2026
From SaaS to intelligence native: the feedback loop.
Intelligence-native systems need agent access to decision artefacts and feedback loops. Why context, not models, is the differentiator — and how MCP, traditional ML, and versioned artefacts fit together.
Apr 16, 2026
Who controls what your AI recommends?
Seasons, campaigns, and weekly events — retail runs on overlapping cycles, and the AI recommender has to keep up with all of them. Notes on the business-rules control surface that lets merchandising teams steer a conversational recommender without editing prompts, filing tickets, or waiting for a deploy.
Apr 13, 2026
Ce que le commerce agentique exige vraiment en production
Macy's annonce que les clients utilisant son assistant IA dépensent 4,75x plus. Sephora vient de lancer une app dans ChatGPT. Zalando déploie son assistant dans 25 marchés. La question pour tous les autres retailers n'est plus 'faut-il le faire ?' mais 'comment l'architecturer pour que ça tienne en production ?'
How it starts
Diagnostic first, build second.
01 · Decision System Diagnostic · From €12,000
Three months, one critical decision. We establish what the system must decide, test what has to be deterministic and what genuinely needs a model, then define the controls required for production. You get evidence, a target architecture, and a costed roadmap.
02 · Design and build
Architecture, the deterministic layer, the model integration, evaluation and the deployment. Scoped from the roadmap, priced on the engagement.
03 · Operate and hand over
Monitoring, evaluation, documentation, and the training to run it without me.