Projects

Work that shows how I think about systems

Selected engineering work, grouped by context: professional production case studies, a personal AI infrastructure lab, and academic research.

Professional work

Production engineering, in a production context

Professional

Agentic Security & Compliance Platform

Problem

Security analysis is slow to generalize: rules and manual review don't scale across repositories, and findings rarely come with structured, actionable remediation guidance.

What I worked on

Agentic workflows that ingest repositories and relevant context, run multi-step analysis, and emit structured outputs — with deterministic policy / guardrail evaluation, risk and impact analysis, remediation workflows, and observability / failure handling.

PythonLLM agentsStructured outputsGuardrails / policy checksRisk & remediation

Read the case study →

Professional

Distributed Query Platform Engineering

Problem

Fleet-scale interactive analytics has to stay predictable: production Trino / Starburst clusters across AWS and on-prem, one federated data layer, and a heterogeneous client / data ecosystem to keep working together.

What I worked on

Distributed query platform engineering across the estate: distributed and federated data access, platform reliability, infrastructure automation, CI/CD, and identity / access integration spanning a heterogeneous data ecosystem.

Trino / StarburstAWS + on-premCI/CDFederated data access

Read the case study →

Personal engineering

Personal lab · hands-on AI infrastructure

Independent of professional work: a personal heterogeneous inference environment, engineered end to end.

Personal lab

Multi-GPU AI Infrastructure Lab

Problem

Heterogeneous GPU systems have different memory, architecture, inference-engine, quantization, and context constraints, so there is no one-size-fits-all serving configuration.

What I worked on

An engineered, containerized heterogeneous inference environment: serving LLMs across different GPUs behind OpenAI-compatible APIs, with per-node engine and quantization selection, long-context budgets, and Tailscale networking for remote access.

SGLangvLLMllama.cppNVFP4 · FP8 · GGUF quantizationLong-context inference

Read the case study →

Research

Academic / research

A forecasting and capacity-planning study — not a production system.

Academic / ResearchR² = 0.86

Predicting Peak Task Periods in Distributed Clusters

Problem

Cluster operators need capacity planning: forecasting when a distributed cluster's load will peak so capacity can be planned ahead of time, not after the fact.

What I worked on

Time-series forecasting of peak task periods using Random Forest and Prophet, with time-aware evaluation so model quality is reported the way it would actually be used.

PythonRandom ForestProphetTime-aware evaluation

Read the case study →