Data & AI leader in ASEAN. Sixteen years building enterprise AI across banking, telecom, energy, and government in Southeast Asia. I build the AI systems and defend the business case for them.
✍️ I write at satsawat.ai: architecture teardowns, strategy frameworks, and honest accounts of what enterprise AI actually costs and returns.
📬 Newsletter: AI in Practice, one enterprise AI teardown a month.
- Why Your AI Agent Will Fail: what the 11% who ship agents to production do differently · Amazon →
- REPLACED: where your career value lives when AI does the 80% · Amazon →
Studies. Does it actually work in production, and what does it really cost?
- tabular-fm-scorecard: do tabular foundation models replace gradient-boosted trees? Four real datasets, cost measured beside accuracy. As accurate, and millions of times more expensive to serve.
- agent-failure-lab: watch compound error kill your AI agent, then watch the mitigations save it. Runnable math, a notebook, and a live calculator.
- sovereign-rag: a paired, fully self-hosted RAG evaluation. BM25 against dense against GraphRAG, six corpora, five languages, and the escape-hatch ablation that moved one model forty points. Every number recomputes from committed artifacts.
- agent-report-card: a performance review for your RAG bot. One CLI, one YAML test file, one markdown verdict a stakeholder can read, with a local judge that reports its own error rate.
pip install agent-report-card· PyPI - tsfm-bakeoff: 13 time-series foundation models against classical baselines, in four isolated environments. Foundation models won 30 of 38 contests.
- llm-inference-arithmetic: the arithmetic behind LLM inference. Roofline, KV cache, Amdahl, and evaluation statistics, as a library and a CLI with no dependencies.
- did-it-land: know whether a side-effect landed, and how to reverse it. Per-vendor effect capsules for durable and saga workflows.
Tutorials. Fundamentals, built by hand and offline-reproducible.
- markov-and-hmm: one matrix prices a loan book, rates an AI agent, and values a customer. Five beginner-friendly notebooks.
- fft-seasonality: build the FFT by hand, prove it on the Keeling curve, then turn hourly demand into a staffing baseline and an anomaly alarm.