class ChenEeHeng:
    """The short version. The long version is on the blog."""

    builds = [
        "LLM systems that turn regulation into structured, reviewable output",
        "agentic pipelines with a deterministic decision layer underneath",
        "the MLOps plumbing that makes both reproducible",
    ]
    also_builds = ["backends", "frontends", "the infrastructure in between"]
    cares_about = ["auditability", "evaluation", "knowing why the system said yes"]
    philosophy = "follow a problem end to end — from the model to the metal"

    def __repr__(self) -> str:
        return "still building, still writing it all down ↓"

Policy-driven risk classification for pull requests. An LLM classifies and scores the change; a deterministic policy table makes the decision. The split matters — the judgement is probabilistic, the ruling is not, and the evidence trail survives review. Python · LLM classification · policy engine

Detection, classification and redaction of personal data, built around GDPR/DSGVO and KYC data-handling workflows. Python · Microsoft Presidio

Multi-agent supervision of a live ML model: monitoring, investigation, remediation, and post-mortem generation. Model risk management, applied to production ML. Python · multi-agent · MLOps

Record and replay Model Context Protocol interactions, so agentic systems can be tested deterministically. Published on PyPI. Python · MCP · PyPI

• 🔭 Building: agentic tooling for regulated-domain workflows — classification with an auditable decision layer underneath
• 🌱 Exploring: OpenTelemetry for agent systems, and evaluation harnesses that catch drift before users do

→ I write about all of it

Engineering notes, machine-learning deep dives, and the things I build along the way.

_{Open to interesting problems and the right opportunity — a note on any of the above reaches me.}