Learning

Studying in public

A public study log, honest about being one: what I am actively learning, what I have already tested in real projects, and what is still planned. The through-line is deliberate practice — fundamentals before frameworks, evaluation before claims, and critical review of anything my AI tools produce before it ships. The reading runs on a written method I now lead a professional circle on: brief a book before page one, recall each chapter closed-book, ship an explanation or a one-week experiment in real work, then review at 48 hours, 7 days, and 30 days.

AI & Machine Learning

Building from both ends: classic ML fundamentals in Python and C++ for depth, and fine-tuning real speech models for Khmer for practice — low-resource constraints punish cargo-cult ML quickly.

  • Classic machine-learning fundamentals — supervised models, evaluation discipline, and the Python and C++ underneath them
  • Speech-model fine-tuning and feature pipelines (Whisper-family models)
  • Evaluation design for Khmer text, where character error rate beats word error rate
  • Error analysis by speaker, device, and noise condition
  • Writing model cards that state dataset provenance and failure modes

Software & Product Systems

Learning to design systems before writing them, ship small typed and tested slices, and lead AI coding agents the way a reviewer leads a team — critically, with explicit approval gates.

  • AI-native development workflows — leading coding agents, reviewing their output critically, and keeping human approval gates explicit
  • System design — decomposing a product into components, data flows, and failure modes before writing code
  • Deterministic pipelines with unit tests and CI
  • Typed content architectures and schema validation
  • Static-site engineering, deployment workflows, and rollback thinking

Business & Strategy

Learning to test commercial assumptions in the field before writing code — and to tell the story afterward, through digital marketing and bilingual technical content.

  • International business coursework
  • Digital marketing and technical content — including short-form video editing in CapCut for English/Khmer developer content
  • Unit economics of agricultural coordination (spoilage, payment timing, fulfillment)
  • Buyer-first validation instead of marketplace-first assumptions

Governance & Commercial Rules

Learning how contracts, terms, and policies decide what AI systems may do — using research agents to survey the primary literature, as a builder who reads the rules, not as a lawyer.

  • Using AI research agents to survey technology-law papers and pull out what actually applies to my products
  • Contracts and terms of service for software and AI products
  • Data licensing, consent, and privacy for speech datasets
  • Risk assessment and human-approval design patterns

Experiments

Each experiment is a falsifiable question carrying an honest state — some have full case studies.

  • Khmer ASR fine-tuning

    Can a fully public Whisper-small pipeline — code, splits, and released weights — reach a usable Khmer character error rate with limited public data?

    Demonstrated Case study →
  • Auditable decision engine

    Can farm-lot assembly be made fully auditable with deterministic code, designed failure cases, and human approval gates?

  • Buyer-first coordination

    Can one recurring buyer order coordinate 30–100 smallholder commitments before harvest?

  • Governance review practice

    Can a structured pre-launch review (data use, terms, oversight, risk) be demonstrated well enough to practice on real products?

    Planned

Reading notes

  • Surrounded by Idiots — Thomas Erikson

    Reading now

    A four-style model of how people communicate. I am reading it to get better at explaining technical work to non-technical people — farmers, teachers, teammates — in the way each of them actually hears it.

  • Master Thinking

    Next up

    Queued as the next read: sharpening the critical-thinking habits I lean on when I review my own AI agents' output instead of trusting it blindly.

Roadmap

Three horizons, drawn honestly: the further out the plan, the fainter the ink. Every item still carries its own label.

  1. Now

    2026 · this term
    • Publish a model card for Kaskor ASR with dataset composition and known failure modes.

      Planned
    • Work through classic machine-learning fundamentals — supervised models and evaluation — alongside Python and C++ practice.

      Planned
    • Practice digital marketing and short-form technical video (CapCut) through English/Khmer developer content.

      Planned
    • Run one recurring buyer pilot for Chomkar OrderLoop, recording every commitment and exception, so the coordination assumptions meet a real order.

      Planned
    • Deepen contracts and terms-of-service study through international business coursework, using AI research agents to survey essential technology-law papers.

      Planned
  2. Next

    toward 2027
    • Take the machine-learning fundamentals past supervised models — implementing the core algorithms in Python and C++ myself before reaching for a library, so the evaluation decisions I make in Kaskor ASR rest on something I have built at least once.

      Planned
    • Practise system design on something larger than a static site: one service with a database, authentication, and a rollback plan — moving from shipping pages to shipping systems.

      Planned
    • Turn the demonstration governance review into a real review of one CHNAI LAB product before its release.

      Planned
  3. Later

    beyond
    • Publish evaluation results by speaker, province, device, and noise condition for Khmer ASR.

      Planned
    • Write publicly about how contracts and policies shaped a real product decision I was part of.

      Planned