AI agent and LLM product development
Most AI features stall between a good demo and a product people rely on. The gap is engineering: routing each task to the right model, grounding answers in your own data, checking output before a user sees it, and keeping the token bill predictable.
I build AI into real products, from assistants and RAG over company data to multi-step agents and decision engines that pair model output with deterministic checks. Recent work includes an EU hosted support assistant that cut one client's support inquiries by 92%, a GPT-4o HR assistant inside a multi-tenant SaaS, and a vision model that turns a photo of a meal into logged nutrition.
What you get
- Assistants and RAG over your documents, database or help center
- Multi-step agents and background jobs that survive long runs
- Model routing across Claude, OpenAI, Gemini and Mistral, with prompt caching and token budgets
- Output checks, human handoff and audit logs where answers carry risk
- Vision and voice features inside web and mobile apps
Relevant work
Moustache AI ↗
EU-hosted, GDPR-compliant conversational AI for 24/7 customer engagement: quick-reply auto buttons, continuous-learning feedback loop, sentiment-based escalation, and an AI ROI dashboard. Cut a client's support inquiries by 92%.
VeltoHR ↗
Enterprise HR & onboarding SaaS for SMBs. Multi-tenant PostgreSQL architecture, employee records, compliance tracking, and workforce analytics. Rebuilt from a prototype into a production platform.
Health AI: Cal & Fit Tracker
AI food-scanning nutrition tracker: photograph a meal and a vision model returns calories, protein, carbs and fat, logged against personalised daily targets. Macro rings, weight milestones, weekly trend analytics, and Apple Health + Fitbit activity sync.
Perzimo ↗
AI fitness coach for women training through midlife and perimenopause: every workout re-planned around sleep, energy, joint limits, and the time actually available. Coaching methodology lives in expert-reviewed frameworks the AI adapts rather than improvises. Voice-guided sessions, adaptive deloads, and Apple Health / Health Connect sync.
Geopolitical Intelligence Platform
Geopolitical decision-intelligence SaaS: plain-language questions, AI scenario simulation with Claude and OpenAI, live market data, and automated PDF and PPT reports, run as queued background jobs on AWS ECS.
Decision Intelligence SaaS
AI-driven decision branching over warehouse data, with row-level security and usage-based billing.
Read the Moustache AI case study →
Questions
Which models do you work with?
Claude, OpenAI, Gemini and Mistral, chosen per task. Cheap, fast models handle routing and extraction, stronger ones handle reasoning, and the choice is written down so you can change it later.
Can the AI use our own data without leaking it?
Yes. Retrieval runs over your data inside your own cloud account, each tenant only sees its own records, and EU data can stay in EU hosting.
How do you keep AI costs under control?
Prompt caching, smaller models where they are good enough, token budgets per request and usage dashboards, so cost is visible before it surprises anyone.