Stack

The shared feature store our ML services read from. Rewrote it from Python to Rust with HTTP and gRPC, halving latency and making it self-service, so it could scale to wide adoption.

~8 ms

p95 feature reads

2x

faster p99 vs Python

Backpressure

safe writes

• Rust
• gRPC
• Tokio
• Bigtable
• Snowflake
• Kafka

A self-hosted Rust memory service for coding agents and local agents, built because existing OSS memory tools did not fit my namespace and control needs.

HTTP + MCP

one service for apps and agents

4 signals

hybrid recall fused at query time

~25 ms

p50 scoped recall, hybrid + rank

• Rust
• MCP
• Axum
• SQLite FTS5
• Qdrant
• ONNX

The trace and eval plane that turns LLM apps from a black box into something teams can trust. I self-hosted Langfuse on a fault-tolerant ClickHouse cluster, with CI-style quality gates for agents.

HA cluster

3 replicas + standalone Keeper quorum

Zero-config

agents self-register for alerting

CI-gated

agent quality checked before prod

• Langfuse
• ClickHouse
• ClickHouse Keeper
• Postgres
• Valkey
• GCS
• FastAPI
• Promptfoo

The single door teams use to reach LLMs. I helped grow it from a solid proxy into a platform-grade gateway with audit logging, guardrails, budgets, batch, and multi-modal support.

Batch API

framework-agnostic

Multi-modal

text, image, audio

Every call

logged + audited

• LiteLLM
• Python
• Kafka
• Snowflake
• GCS

The ML lifecycle and monitoring stack our teams build on. I own and harden our internal builds of OpsML and Scouter and wire them into the platform.

Scouter

stood up end to end

Kafka + gRPC

monitoring pipeline

Rust core

internal fork owner

• Rust
• PyO3
• OpenTelemetry
• Kafka
• Postgres

The service that put surge pay on autopilot for last mile delivery. A realtime, market-aware Keras model served async over FastAPI, built on the learnings of an earlier offline version.

1st

surge model trusted to automate pay

Realtime

market-aware serving, async over FastAPI

Owned E2E

training through production serving

• Keras + JAX
• Polars
• FastAPI
• Feature store
• GPU

The one registry the whole company tracks experiments and ships models on. I run it as production infrastructure, backed by Cloud SQL and GCS, and make the operational calls that keep it boring.

• MLflow
• Cloud SQL
• GCS
• Datadog
• Prometheus
• Harness

A self-hosted Rust remote-dev stack. Lobsterbox leases Docker or Hetzner runners, and lobsterfleet drives them from the browser with terminals, VNC desktops, and Codex-ready boxes.

Docker + Hetzner

leases from one CLI

Terminal + VNC

in the browser

Agent API

drive boxes headless

• Rust
• Axum
• Yew/WASM
• SQLite
• Hetzner
• Docker

Self-hosted speech to text in Rust on whisper.cpp. Runs Whisper on our own boxes, faster and lighter than the Python version, with no per-minute bill and no audio leaving our walls.

~3-4x faster

CPU inference vs Python

~2x less RAM

vs PyTorch Whisper

~3x cheaper

compute to run vs Python

• Rust
• whisper.cpp
• Symphonia
• Axum
• Tokio

A self-hosted experimentation platform where I try bleeding-edge tools and new methodologies on real infrastructure, from the network edge to GPU compute.

• Self-hosted
• Network segmentation
• Clustered compute
• SSO
• LLM gateway
• Remora

Helped launch geocode correction for Target Last Mile by showing corrected pins matched delivery evidence and sizing the annual orders exposed to bad source coordinates.

Launch-validated

corrected pins matched delivery evidence

PoD-backed

checked against proof-of-delivery locations

Impact-sized

annual order exposure from bad source pins

• Geospatial analysis
• PoD validation
• SQL
• Launch analytics
• Target Last Mile

Validated the 20-minute dispatch window launch with geo exits that separated real departure from early pickup reports.

Geo-validated

launch window confirmed on cleaned signal

Tighter geofence

separated real departure from load time

Signal QA

filtered unreliable geo sessions

• Geo signals
• Dispatch analytics
• Geofencing
• Outlier filtering
• TLMD

Rebuilt the package detection behind proof of delivery: off YOLO onto a recall-first Mask R-CNN, with a SAM2-powered Label Studio pipeline for labeling and evaluation.

+7 pts

recall

+4 pts

precision

~60%

missed packages recovered

• Mask R-CNN
• PyTorch
• SAM2
• Label Studio
• FastAPI
• GPU

The first visual operations platform for Target Last Mile. It gave operators one place to see pay, store coverage, delivery surge, and order density across markets.

1st

visual ops platform

12 SCs

daily ops use

4+

critical ops views

• Python
• Streamlit
• Plotly
• Snowflake
• H3
• Redis
• GCP

A production route operations assistant I designed and built end to end, with 8 validated tool actions and full Datadog and Langfuse traces.

Production

agentic system for route operations

8 tools

validated, guarded actions

3 protocols

AG-UI, A2A, MCP

• Google ADK
• AG-UI
• A2A
• MCP
• Langfuse
• Datadog

The durable memory and task layer for our agent stack. A Rust service that speaks gRPC and HTTP over Postgres, with Python clients that give Google ADK, A2A, and LangGraph agents drop-in durable sessions and tasks.

gRPC + HTTP

dual transport, one service

Drop-in

durable agent state in one call

~15 ms

p95 gRPC read

• Rust
• gRPC
• Protobuf
• Postgres
• Python
• Google ADK

Research Assistant

Research assistant work at Ohio State: a Python scheduling app and operations research model for assigning incoming trailers to strip doors in a cross-dock.

Python

trailer assignment app

LP

strip-door scheduling model

DES

workforce and throughput simulation

• Python
• Linear programming
• Discrete-event simulation
• CPLEX

Research Assistant

A Markov chain Monte Carlo model for wind turbine maintenance that tested time-dependent failure rates and showed a 20% maintenance cost reduction in simulation.

20%

modeled maintenance cost reduction

MCMC

maintenance strategy simulation

Time-based

failure-rate modeling

• Python
• Markov chains
• MCMC
• Cost modeling

Simulation project

A discrete event simulation of an Ohio State cafe that tested self-ordering kiosks and workforce changes, showing a 16% modeled wait-time reduction and at least $288/day in savings.

16%

modeled wait-time reduction

$288/day

estimated daily savings

• Arena
• Discrete-event simulation
• Queueing
• Workforce analysis