Case studies
Selected work across AI-native engineering, research infrastructure, and production pilots.
Production LLM gateway + cost controls
Designed and deployed a multi-model LLM gateway with TensorZero on Fly.io for a team that needed predictable latency, cost limits, and A/B testing across model providers.
Challenge
Token spend spiking; no visibility into which model or prompt was driving cost.
Approach
Multi-model gateway, per-function budgets, eval harness, and observability hooks.
Outcome
Cost predictability, latency SLOs, and the ability to swap models without code changes.
AI roadmap for a seed-stage SaaS startup
Helped a seed-stage team go from "we should use AI" to a scoped 90-day pilot with clear evals, model choices, and a hiring plan.
Challenge
Founders had AI ideas but no ML engineer and no clear first milestone.
Approach
Discovery workshop, use-case scoring matrix, stack recommendation, and pilot scope.
Outcome
90-day plan with $15k pilot budget and clear go/no-go criteria.
AI research copilot for an investment newsletter
Building a 4-week pilot to ingest earnings-call transcripts and SEC filings into a cited Q&A interface so analysts spend less time reading and more time judging.
Challenge
Research team drowning in filings, transcripts, and news feeds; zero AI expertise in-house.
Approach
4-week pilot: RAG pipeline, citation tracking, theme detection, and human review gates.
Outcome
Pilot scoped for $8,500 with optional $1,999/mo retainer for expansion.
Product proof points
Public product projects
These are public projects I can talk about. Commercial work is under NDA, but references and architecture reviews are available after an intro call.
VCA Inference Cache
Problem: High-throughput LLM serving workloads waste compute on repeated prompt prefix processing.
Approach: Built a key-value cache layer optimized for inference workloads, with eviction policies tuned for transformer attention patterns.
Outcome: Reduced redundant computation for repeated prompt structures; cache hit rates scale with conversation length and batch size.
Apollo Mission Simulator
Problem: Historical space-mission simulations are either toy models or closed-source.
Approach: Combined a 3D Apollo command module with the real yaAGC guidance computer and Keplerian orbital mechanics in the browser.
Outcome: A runnable, inspectable simulation that proves we can bind real embedded systems, physics, and 3D rendering in a single web app.
Merlin LLM Router
Problem: Model selection is usually static, so latency and cost suffer on heterogeneous request traffic.
Approach: Built a multi-provider router that picks the cheapest capable model per request using sub-millisecond reinforcement-learning scoring.
Outcome: Lower per-request cost without sacrificing quality; routing decisions improve as request patterns change.
DGX Spark Memory Planner
Problem: Developers waste hours tuning memory and quantization settings for NVIDIA DGX Spark workloads.
Approach: Created a unified memory-budgeting and quantization advisor that models batch size, sequence length, and precision trade-offs.
Outcome: Faster, less error-prone model sizing; fewer out-of-memory failures during experimentation.
Have a project that belongs here?
Tell us what you're building. We'll reply within one business day with a scoped plan and estimate.
Book a strategy call