AI engineering that works in production

I build AI features that do real work: answering questions from your own documents, running multi-step tasks with agents, and adding intelligence to products people already use. Each one ships with an evaluation set, so we both know how well it works before anyone relies on it.

What I build

• Retrieval (RAG) assistants that answer from your documents and cite where the answer came from.
• Agents and agent teams that carry out multi-step work with clear limits and human checkpoints.
• AI features inside existing products: search, drafting, classification, summaries.
• Evaluation and monitoring, so quality is measured, not assumed.

How I approach it

1. Start with the job, not the model. What task, for whom, and what does "good enough" look like?
2. Pick the simplest thing that works. Often that's better prompts or retrieval before any fine-tuning. See RAG, fine-tuning or better prompts?
3. Measure it. A small, honest test set from real examples, run on every change.
4. Design for trust. People should see sources, understand limits and stay in control.

Proof

• Z-web Blueprint Agent: an agent that crawls a website and explains its architecture. It ran the 43-site research behind this website.
• e-Serbisyo and PasinPay: both designed and built by teams of specialised AI agents that I set up and directed.
• Site X-ray: try the blueprint agent's approach in the Lab.

Frequently asked

Do we need a custom model? Rarely. Most business problems are solved with an existing model plus your data and good design. I'll tell you if yours is the exception.

How is our data handled? Only what the feature needs, processed where you agree, never used to train public models without your consent. Details in Principles.

How do you know it's working? We agree on success criteria first, then test against real examples before launch and keep testing after.

Start a conversation