azamzam.aiI build agentic RAG systems for high-stakes knowledge work.
Private, accurate, citation-grounded AI assistants — built on your licensed content, deployed in your environment. For teams in legal practice, compliance, and regulated certification work.
What I build
Generic LLMs are a poor fit for regulated work. They hallucinate, they cannot cite, and the moment you feed them confidential content you have a data-handling problem rather than a knowledge problem. The off-the-shelf chatbot that works fine for marketing copy will not survive contact with a partner reviewing its first piece of legal output.
I build purpose-engineered RAG systems on the client's own corpus — statutes, contracts, internal policies, audit manuals, case files. The work is in the parts that don't show up in a demo: chunking that respects the structure of legal text, retrieval that surfaces the right section across cross-references and amendments, an agentic layer that knows when to look again and when to refuse, and an evaluation harness that lets the team measure regressions before users see them.
I work with teams whose knowledge base is the regulatory text itself — law firms, certification auditors, financial services compliance, sustainability frameworks.
How I think
Three of six design principles, earned from shipping.
Retrieval is a search problem before it is an ML problem.
Most RAG failures come from bad chunking and missing metadata, not bad embeddings. Get the corpus shape right and a mid-tier model outperforms a frontier model fed sloppy chunks.
Every answer must cite; every citation must verify.
In regulated domains, an unverifiable claim is worse than no claim. Citations are not decoration — they are the trust contract between the system and the user.
Evaluation comes before optimisation.
You cannot improve what you cannot measure. The eval harness is built on day one, not month three. Every change is a measured delta.
Featured case study
Sarawak Labour Law — agentic RAG over Cap. 76 and Act A1754
A worked example of building citation-grounded legal AI for a complex jurisdictional domain. The Sarawak Labour Ordinance, its 2022 amendments, and the interplay with federal labour law form a corpus where naive retrieval produces confidently wrong answers. The case study walks through the chunking strategy, the embedding choice, the agent loop, and the evaluation results — honest about what works and what doesn't.
How I work
A repeatable methodology, eight steps. Three of them previewed below.
- Domain discovery. Three to five interviews to identify the decisions the system must support — not the questions the buyer thinks it should answer.
- Corpus analysis. Read the corpus. Map structure, citation patterns, and ambiguity hotspots before any architecture decision.
- Evaluation harness. Built in week one. Every change gets a measured delta against the same test set.
About
Twenty-five years writing software, most of it Java backends and the kind of long-lived systems that have to keep being right. Based in Malaysia.
A career spent shipping production SaaS in regulated and operational domains gave me the instincts this work depends on: clean domain models, transactional thinking, and a low tolerance for systems that look right but aren't. Agentic RAG is what I do now because it is the place where engineering rigour matters most and is most often skipped.
Building something in a regulated domain?
If you are evaluating whether AI can help your firm or team handle the kind of content that ChatGPT cannot reliably reason over, I'd like to hear about it.