AI Engineer @ NeuralMetrics · Python · FastAPI · LLM systems
Building CHRIS, an autonomous premium-audit agent for US commercial insurance.
I build AI systems that replace work people shouldn't have to do by hand.
- 6+ years shipping production Python — now building autonomous agents end to end.
- At NeuralMetrics: CHRIS validates audit data, detects the classification errors that move premium, adapts to compliance rules that shift by state and year, and analyzes risk exposure.
- Day-to-day: FastAPI, PostgreSQL, the Claude API, Model Context Protocol, and AWS (Lambda, Bedrock, ECS/ECR, S3, RDS).
- Interested in agent architectures, MoE inference systems, applied NLP, and InsurTech.
Quantization as Cache Amplification: Trillion-Parameter Mixture-of-Experts Inference on a Commodity Laptop 📄 Paper (Zenodo, DOI 10.5281/zenodo.21856981) · 📊 Traces, code & artefacts (Hugging Face)
Weight quantization is usually sold as footprint reduction. This paper shows that for
offloaded Mixture-of-Experts inference the real lever is cache capacity: an LRU expert
cache hits exactly 0% whenever it holds fewer than the k·L expert slots a single
token touches — a phase transition measured on real routing traces — and quantization is
what carries a system across it.
| Result | Value |
|---|---|
| LRU hit rate below per-token working set | 0.0% (all capacities tested) |
| Popularity-pinned cache @ 10% capacity | 22.9% (vs 0.0% for LRU) |
| Frequency-conditioned allocation @ 1.51 bits | 22.02 PPL vs 25.54 uniform (13.8% better at identical rate) |
| Sub-2-bit codec @ 2.01 bits, WikiText-2 | 12.17 PPL (bf16 reference: 8.11) |
Everything was measured on one laptop (RTX A500, 4 GB VRAM; 32 GB DRAM; consumer NVMe) — sub-2-bit codec, raw OLMoE-1B-7B routing traces, and every measurement artefact behind the numbers.
| Project | What it is |
|---|---|
| ssh-mcp | My own MCP server for remote development over SSH — files, commands, search, git, and binary-safe SFTP transfer via Paramiko. Cross-platform, Python. |
| CHRIS (private, @ NeuralMetrics) | Autonomous premium-audit agent for US commercial insurance: data validation, classification-error detection, dynamic regulatory compliance, risk-exposure analysis. |
| quantization-cache-amplification | Codec, routing traces and benchmarks behind the MoE paper above. |
- OpenHands (formerly OpenDevin) — #3285 merged: clear history at the start of a new task.
- LiteLLM — #10548 merged: Gemini 2.5 Pro max_tokens fix · #17298: proposed Claude Code provider integration.
- freqtrade — merged: #6545 partial exit using average price, #6540, #6432.
- NostalgiaForInfinity — 7 merged PRs on leverage handling, trailing stops and backtesting.
- openai/codex — running my own fork and taking part in issue discussions.
Anthropic — Building with the Claude API · Claude Code in Action · Model Context Protocol: Advanced Topics · Introduction to agent skills · Introduction to subagents · Claude with Amazon Bedrock · Claude on Google Cloud
Cisco — Python Essentials 1 · Python Essentials 2
Panels regenerate daily via lowlighter/metrics.