AI systems that survive contact with real data.
I build retrieval systems, agent platforms and voice pipelines for products that are already running the part most teams skip: making a model behave predictably once it's handling someone's actual business.
Selected work
Three of the harder problems.
Nova ADK
Full-stack platform for building and running AI agents. Intent routing, guardrails with PII detection, trust evaluation on every output, memory, tool calling and multi-agent orchestration.
VectorLead
Every inbound Meta lead called within a minute by a voice agent that qualifies in conversation, scores the result and routes it onward. The calling provider is swappable without touching the pipeline.
Hybrid retrieval engine
Three retrieval channels fused in one query, cross-encoder reranking, citations verified against their sources, and an evaluation harness so quality is a number rather than an opinion.
Selected case studies
More from the portfolio.
Multi-tenant AI services platform
One shared substrate for routing, caching, fallback, metering and audit around twenty-five AI features inherit all of it rather than each wiring its own infrastructure.
Read case study
Enterprise platform AI
AI over a live CRM installation. Metadata crawl into a relationship graph and configuration index, grounded answers, generated artefacts compiled and verified against the real schema before any deployment.
Read case study
Healthcare marketplace search
Patients search by symptom, not specialty. Keyword and semantic retrieval run together and are fused, with a hand-built medical terminology layer expanding the query before either sees it.
Read case study
Commerce discovery & recommendations
Local service discovery where geography constrains rather than competes with relevance, plus a collaborative filtering recommender and a scheduled notification engine with delivery auditing.
Read case study
Market intelligence engine
Sentiment, technical indicators and strategy backtesting computed deterministically, then read by a reasoning model that writes the analysis. The model explains numbers it was given never produces one itself.
Read case study
Generative media pipeline
Script to synthesised speech to rendered video, built around job state, exponential backoff, duplicate prevention and a dead-letter path because generative media APIs fail more than their docs admit.
Read case studyHow I build
Principles that show up in every system.
Measured, not asserted
Golden datasets and ablation runs from week one. A change is proven better not argued about and regressions appear the day they land.
Failure has to be visible
When a stage degrades, the response says so. A system that quietly returns worse results looks identical to one that works.
Preview before it writes
Anything irreversible is rendered as a proposed change first the difference between an agent people trust and one they switch off.
Costs are designed in
Task-appropriate model routing, caching and per-tenant metering from the first release. Unmetered token spend is how a feature dies on margin.
Ready to build something that holds up?
Eight years of engineering. A deep portfolio of production AI systems. No slide decks just working software.
See all the work →