I build ingestion systems, feature pipelines, retrieval infrastructure, and LLM agent evaluation tools for messy real-world domains.
|
I work on practical data and ML systems: pulling noisy signals from fragmented sources, shaping them into queryable datasets, and building evaluation loops that make model behavior easier to inspect. Most of my public work sits in domains where the data is awkward: transaction streams, retrieval networks, industrial telemetry, and tool-using agents. The domain changes; the engineering pattern is consistent: ingestion, validation, feature extraction, structured storage, and reproducible evaluation.
Collaborations on:
|
| Repo | Why it matters |
|---|---|
defi-risk-reasoner |
Tool-using LLM agent with structured risk reports, labeled eval cases, and scoring scripts |
mev-pattern-miner |
Event-stream mining pipeline that builds labeled parquet datasets and trains classifiers |
token-risk-feature-store |
Multi-source feature pipeline for token-level risk signals and offline analysis |
industrial-telemetry-ingestion |
Industrial telemetry ingestion and monitoring system using MQTT, Modbus, IEC 104, Express, and Vue |
distributed-retrieval-pipeline |
Distributed retrieval pipeline with schedulers, resolvers, workers, retries, and result aggregation |
verifiable-agent-memory |
Verifiable trace registry for AI agent decisions, with contracts plus a Python SDK |
| Repo | Demo artifact |
|---|---|
industrial-telemetry-ingestion |
MQTT/Modbus/IEC 104 telemetry flow, entity graph, and sample payload |
defi-risk-reasoner |
Tool-using LLM agent eval sequence, scoring flow, and sample metrics |
mev-pattern-miner |
Blocks-to-parquet-to-classifier pipeline and detector metrics |
token-risk-feature-store |
Source adapters, feature entities, parquet snapshots, and sample feature row |
distributed-retrieval-pipeline |
Scheduler/resolver/worker sequence and sample retrieval result |
verified-retrieval-cache |
Merkle proof cache flow and sample window-path artifact |
verifiable-agent-memory |
Agent trace commitment sequence and verification result |
The opinions and code in these repos are mine alone. Most repos are prototype-grade; read each README before assuming production readiness.

