HM / Lab · 2026

Build the system.
Expose the tradeoffs.

The lab is where I turn architecture ideas into inspectable experiments: retrieval pipelines, evaluation loops, scientific interfaces, agents with typed tools, and the infrastructure required to make them trustworthy.

Active notebook 6 experiments
Principle 01 Evidence before fluency.

Every generated answer should be traceable to the context that justified it.

Principle 02 Tools before autonomy.

Give models explicit, typed actions before asking them to behave like open-ended agents.

Principle 03 Evaluation before scale.

Measure groundedness, task success, latency and failure modes before optimizing throughput.

Experiments

Working systems, not screenshots.

These are design studies for the kinds of production problems I care about. Each one isolates a technical decision and makes the failure modes visible.

EXP / 001 architecture

Evidence-first RAG for scientific review

A retrieval architecture for scientific teams where provenance, document hierarchy and abstention matter more than conversational polish.

Docs → Parse → Hybrid retrieve → Evidence → LLM → Review
RAG provenance biotech
EXP / 002 evaluation

Question prediction without data leakage

An evaluation harness for predicting regulator-style questions while separating historical question banks from held-out validation data.

Coverage Novelty Grounding
evals retrieval
EXP / 003 agents

Typed-tool agent boundary

A narrow agent pattern where models can read evidence, call approved tools, propose actions and stop at a human approval gate.

Observe Reason Tool Review
agents tool use human-in-loop
EXP / 004 interface

Scientific workflow UX

What changes when an AI interface is designed around evidence inspection, experimental context and decisions—not chat.

UX scientific software decision systems
01 Input evidence 12 sources
02 Model synthesis 3 claims flagged
03 Human review required
04 Action locked
EXP / 005 distributed systems

Durable state for agentic workflows

A reference pattern for retries, idempotency, partial failure, tool execution and human escalation when probabilistic planners call deterministic services.

Plan Validate Execute Persist
agentic AI distributed systems idempotency
EXP / 006 security

Authorization-preserving retrieval

An enterprise RAG boundary where source-system permissions participate in retrieval before evidence reaches the model context or tool layer.

Identity → Policy → Retrieve → Evidence
enterprise RAG authorization MCP tools
What I’m testing next

Small models, strong systems.

How far can a compact model go when retrieval, tool interfaces, domain context, caching and evaluation are engineered properly? That question is more interesting to me than simply moving to a larger model.