I build the system that puts AI to work across a chip-design company: governed agents on every engineer's machine, knowledge siloed by project, one gateway for every model, and a ledger that ties each customer requirement to the work and evidence behind it. At Aion Silicon, after three years designing chips.
Inside the experience vault: the graph, retrieval and checks
Before you build is the cheap window. I lay everything out up front (the plan, the workflows, the system) while being wrong still costs an edit, and everyone can see it and push back. Once a project ramps up, every mistake costs more and the team ends up firefighting.
A system should be block-diagram-able, and the code should mirror the diagram. When two approaches tie, I pick the one that’s easier to draw.
It’s the absence of reporting. No evidence means unknown, never green, and unknown stays distinct from zero.
A wrong choice is fine if the module can be replaced. Tight coupling is the real mistake: it turns a wrong choice into a rewrite.
An LLM is a probability machine, so most of a workflow should be plain, testable code. The model takes the judgement calls, and security is enforced outside it.
Work counts when the check passes. A test only counts once I’ve seen it fail, and a pass over zero items is a failure.
What one project learns becomes a rule every agent loads, and every new project starts from the last one’s blueprint: cut what you don’t need, change what differs, add what’s missing.
| Modules with typed ports | Schemas and typed fields before any code |
| Architecture exploration before RTL | Lay the whole project out before building any of it |
| A block with no regression results isn’t passing | A task with no evidence is unknown, never green |
| A golden reference model | An independent ground truth behind every eval and grader |
| Re-verify after a netlist change | Evidence hashed to its inputs: change one, re-check everything downstream |
| A derivative chip starts from the last flow | Every project starts from a template, and its lessons go back in |
Three problems from the last year, and what changed once I looked past the first answer.
Hashing every record, like a blockchain, proves nothing was changed after it was written. It doesn’t prove the record was true. And one writer checking itself is the opposite of what makes a blockchain trustworthy.
Designed Jarvis around independent checkers and a mirror of the ledger the writer can’t rewrite. Work done outside the system gets recorded, but it can’t turn a task green until it’s checked inside.
The instinct is to blame the model. Measuring first showed 18% of the corpus had silently converted to empty text, and the reranker added to improve results was cutting nDCG by 39%.
Fixed the conversion and made it provable, which became doc2md’s lossless gate. From then on, a change stayed only if the benchmark backed it.
Rules written into a prompt aren’t controls. A model can be talked out of an instruction; a hook outside the model can’t be.
Enforced the policy outside the model: IT-managed rules users can’t override, hooks that block secrets, every action logged per person. Then wrote down what isn’t covered yet, so IT never claims more than the system delivers.
Building the company AI platform, largely as its sole engineer: the gateway, retrieval, agents, evaluation and governance that let the engineering organisation use AI safely.
Re-engineered a RISC-V vector core to drive an external VPU accelerator, plus flow automation in Python, Tcl and Bash.
Front-end RTL across IP and SoC projects; set up the company-wide lint methodology; owned DFT on an automotive SoC.
First-Class Honours. Malaysia campus, then UK.
The agentic ledger that grounds a chip project in real tasks. A customer’s sentence becomes goals, requirements and tasks; each task gets its own workspace, and agents do the work with the existing EDA tools.
Every output is tied to its inputs by fingerprint, so a changed requirement or file re-checks everything built on it. Projects start from a template with lessons from past ones built in, and managers, engineers and agents all read the same ledger.
Claude Code for a shared Linux engineering fleet, run by IT: a policy users can’t override, hooks that block secrets and log every action per person, a check-first installer, and daily, drift and weekly reports.
Bash · Python · systemdThe company’s Bedrock setup as code: the model registry, an IAM guardrail with tagged profiles and budgets, a LiteLLM gateway with keys for AgentCore and LangGraph agents, and a switch that runs Claude Code on Bedrock.
Python · LiteLLM · AWSHow lessons travel between my projects. Claude sessions and design docs become a linked Obsidian vault; lessons become rule decks, and the decks ship as skills that agents in every other repo load.
Python · Obsidian · Claude CodeTurns Office documents into token-efficient Markdown for RAG, with a second pass that proves nothing was lost: docx, xlsx and pptx must reach full token recall, and PDFs report exactly what they dropped.
Python · docling · OOXMLThe skeleton my repos start from, Jarvis and bedrock-platform included. One labelled layout that humans and coding agents share, with gates that fail the build when structure, docs or links drift.
copier · Python · CIWatches for documentation drifting from code, asks a pluggable LLM backend to fix or invalidate it, and logs every verdict for human review against a versioned public schema.
Python · LangGraph · FastAPIHappy to talk agentic systems, LLM platforms, or getting AI past an IT security review.