Work
What I have delivered
Four sectors, four published articles under someone else's masthead, three open-source tools and two books — all of it my own delivery rather than a firm's portfolio. Two customers cannot be named, so the sector and the scale are all that is claimed for them.
Sectors
Government & public sector
The constraint. National-scale citizen services where the constraint is rarely model quality — it is security posture, data sovereignty, and whether the thing can be operated in-policy once it is live.
Architecture and solution leadership for digital government workloads — primarily in Thailand, with regional exposure across ASEAN — spanning document understanding, citizen-facing assistants, and border-control modernisation. Several are published in the open.
- AI sovereignty
- Agentic workflows
- Multimodal document understanding
- RAG over policy corpora
Banking & financial services
The constraint. Where the modelling is rarely the hard part — getting the output into the channels and decisions that actually touch customers is.
Lead data scientist work on data-as-a-product platforms and their integration into customer-facing banking channels, shifting analytics from one-off reports toward reusable products other teams could build on.
- Data-as-a-product
- Retrieval-augmented generation
- Agent guardrails & approval flows
- Model risk & explainability
Telecom & digital media
The constraint. Subscriber bases large enough that a one-percent movement in segmentation accuracy is a material revenue number — and small enough modelling errors compound fast.
Segmentation and analytics leadership at national telecom scale, including founding a new analytics business unit and building micro-segmentation models that fed directly into commercial decisions.
- Micro-segmentation
- Agent evaluation & observability
- Compound failure analysis
- Real-time inference
Energy & industrial
The constraint. Heavy industry, where data is plentiful and messy and the cost of being wrong is measured in plant operations rather than click-through.
Senior data science work on energy-sector analytics.
- Time-series forecasting
- Foundation models for forecasting
- Anomaly detection & triage
- Agent escalation paths
Delivered
Three are described on the AWS Public Sector Blog, which outranks a self-reported figure for anyone reading sceptically.
- A digital arrival card at a national border AWS Public Sector Blog →
- Multimodal document understanding for government workflows AWS Public Sector Blog →
- KMITL’s prospective-student advisor, on Amazon Bedrock AWS Public Sector Blog →
- A national postal operator serving 7M+ users, and a national super-app Not public
Published
On the AWS Public Sector Blog, describing systems that are live.
- Modernizing border control with digital arrival cards on AWS Cloud Read →
- Empowering government document understanding with Amazon Nova Multimodal Embeddings Read →
- KMITL transforms prospective student guidance with Amazon Bedrock Read →
- Empowering bacterial genomics education with Amazon WorkSpaces Read →
Open source
Runnable, not illustrative. Every repository below is public and MIT-licensed.
agent-report-card
An evaluation harness that calibrates the judge before it grades anything: the model doing the scoring sits a fixed answer key first, and its measured error rate is printed in every report. A judge you have not measured is an error rate you have silently inherited. On PyPI at 0.2.0, with CI.
tsfm-bakeoff
Five time-series foundation models against three classical models, a gradient-boosted tree and four naive baselines — thirteen in the field — across ten datasets and four forecast horizons. Foundation models won 30 of 38 contests, and the model topping the public leaderboard won none. Three isolated Python environments, and a scorer that imports no model code.
agent-failure-lab
Compound error in multi-step agents, made runnable — a browser calculator, a zero-install Python simulator, an executed notebook, and a real 8-step agent that runs against a local model and classifies every failure. Companion to the book and to the compound-failure article.
Books
Why Your AI Agent Will Fail
And What the 11% Who Ship to Production Do Differently
Only 11% of enterprise AI agent pilots reach production. A practitioner’s playbook for the decisions — not the demos — that separate the two.
REPLACED
Why I Built the AI That’s Restructuring Your Career and How to Stay Ahead of It
You’ve been good at your job for years. Now AI does the 80% you’ve been running on autopilot — a framework for finding where your career value actually lives.