I build LLM inference, agent tooling, and the systems under them.

Single-GPU LLM Serving Optimization

Independent researchCUDAOngoing

Serving a 27B model at long context on one consumer GPU, and finding the throughput that the stock kernels leave on the table.

System

SGLang serving Qwen3.8-27B (NVFP4) on one RTX 5090: FP8 GEMM dispatch, a fused FP4 activation kernel, NVFP4 autotuning, and a quantized speculative draft model.

Responsibility

Sole owner: kernel work, profiling, draft and checkpoint selection, and the benchmarks behind every number below.

Proof

Throughput up 36% from kernel work; 145 tok/s sustained (+59%) at a 216K-token context. NVFP4 GEMM latency cut 30% (72.9 → 51.1 µs). Extended to 256K context at 131.5 tok/s decode, and 103 aggregate tok/s across four concurrent requests (from 74).

CUDASGLangNVFP4Nsight

Wispr Agent CLI & Evaluation Harness

PersonalGo2026

A Go CLI that gives coding agents bounded, structured access to local meeting summaries, search, and timestamped transcript evidence — so an agent can cite what was actually said without ever getting free rein over the database.

System

Read-only SQLite access, deterministic ID resolution, bounded pagination, structured JSON envelopes, speaker resolution, and explicit untrusted-content boundaries.

Responsibility

Sole owner — design, implementation, the agent-integration skill that defines safe usage patterns, and the evaluation harness.

Proof

Cross-model evaluation over 10 behavioral cases scoring correctness, tool use, latency, tokens, and cost. Findings drove search and agent-guidance changes; all 20 targeted reruns stayed correct and met their per-case tool-call budgets.

GoSQLiteAgent toolingModel evaluation

Klinebot Production Automation Daemon

HyperTrendGo2025—2026

A Go daemon running unattended against live exchange and blockchain APIs. The hard part was never the integration — it was making the thing fail safely when the network, the clock, or the operator does something unexpected.

System

SQLite persistence, deterministic scheduling, live execution, event logging, health and NTP checks, separated fund-control and limited execution keys, and pause/resume safeguards.

Responsibility

Led design and implementation, including application scaffolding and production code, then shipped static binaries and Docker images for AWS VPS deployment.

Proof

Every client who received it deployed it. It ran roughly three months in production with no significant incidents reported.

GoSQLiteDockerAWS

Client work — architecture and outcomes only.

Reproducible Research Data Package

HyperTrendPython2026

Roughly 20 GB of compressed historical data, packaged so that somebody else can re-run the analysis and get the same answer.

System

Zstandard-compressed Parquet, DuckDB query recipes, provenance tracking, quality validation, and explicit coverage semantics.

Responsibility

Built the package and its client-facing technical materials, and was selected to lead interactive instruction on it.

Proof

Shipped as a reproducible analysis workflow rather than a data dump — sources, validation checks, and coverage limits are all inspectable by the reader.

PythonParquetDuckDBZstandard

High-Availability K3s Compute Cluster

Personal infrastructureOngoing

A mixed-hardware K3s cluster I run so that agent and automation workloads have somewhere isolated, persistent, and cheap to live.

System

Multi-node K3s across heterogeneous machines — 50+ CPU cores and 400+ GB aggregate memory — with GPU scheduling, NAS-backed shared storage, and redundant nodes for workload continuity.

Responsibility

Built and operate it: Ansible provisioning, Grafana monitoring, custom CI/CD runners, storage, scheduling, upgrades, and recovery.

Proof

Runs isolated containerized coding-agent and automation workloads daily. Capacity figures are deliberately conservative.

K3sAnsibleGrafanaLinuxDocker