Matthew Goldey

I build production LLM, retrieval, and speech systems — and the evaluation infrastructure that tells you whether they actually work.

Director of Data Science & AI Zionsville, IN GitHub LinkedIn Scholar Substack

Now

I own the AI function at a media technology company serving 100+ enterprise newsrooms, taking greenfield projects through production and setting AI strategy. That means fine-tuning and serving open-weight models on cloud GPUs we run ourselves, building hybrid retrieval, and putting evaluation gates in front of anything that reaches a customer.

Model discipline is my common thread across eighteen years of research and industry work in computational chemistry, speech recognition, and LLMs: measure honestly, test with real data, and solve the real customer pain. Most of what goes wrong in applied ML is an evaluation or user experience problem.

Experience

2024 —

Director of Data Science & AI

Nota · remote
  • Own the entire AI function; set strategy with the CEO and run all production ML
  • Fine-tuned a mid-size model that beat every Gemini frontier model on internal LLM-as-judge evals across seven production templates
  • Cut projected inference spend 80% by moving managed cloud ML to self-hosted vLLM on Kubernetes
  • Reduced hallucination rate 5.9% → 1.2% with targeted evaluation gates
  • Shipped DRAFT, an LLM article system: 42K+ articles across 100+ stations at 4/5-star satisfaction
  • Added RAG citations and knowledge injection to improve editor trust in AI-generated draft articles
  • Built production hybrid search from components (Postgres full-text, trigram, and vector retrieval with a re-ranking stage); cut irrelevant results 44% → 8% on test sets, and latency from ~10s to under 100ms
sources retrieval + citations 20B+ model self-hosted eval gate per-criterion ✓ fails → regenerate, never ships hallucination rate 5.9% → 1.2%
Nothing reaches a newsroom without passing the gate. The gate is the product.
2021 – 2024

Director of Data Science → Principal

Truleo · remote
  • Led ML for a police body-camera analytics platform, then owned platform-wide ML architecture; managed the data science team
  • Built an inference pipeline processing thousands of audio hours daily at 60× lower cost than competitors
  • Reduced ASR word error rate 25% and speaker diarization errors 33%
  • Shipped an LLM-powered phone-call-to-report product adopted by multiple police departments
thousands of audio-hours / day voice detection ASR WER −25% diarization errors −33% intent + sentiment 60× cheaper / hr whole pipeline
Volume is the constraint: at thousands of hours a day, cost per hour decides whether the product exists. Pipeline after US 12,014,750, FIG. 1 — named inventor.
2017 – 2021

Data Scientist → Director of Data Science

Green Key Technologies · Chicago
  • Led data science for NLP and speech in financial services; managed two data scientists who developed customer-specific NLP solutions
  • Led embedded engagements at Bank of America Merrill Lynch and RBC, owning scoping, delivery, and the client relationship
  • Reduced ASR errors 40% on noisy trading-floor audio; validated on-premise deployments across five countries
bank security perimeter — data never leaves desk audio + vocabulary fine-tuning runs in-place adapted model WER −40% 5 countries only the model crosses the boundary
The constraint was the product: a bank that cannot export audio can still get a model tuned to its own desks.
2008 – 2017

Research — computational chemistry

UC Berkeley (PhD) · University of Chicago (postdoc)
  • Developed electronic structure methods reducing model errors 80% and compute cost by orders of magnitude
  • Led a three-person software team across three institutions; won 4M CPU-hours competitively
Physical Chemistry Chemical Physics front cover: a fullerene acceptor surrounded by donor polymer chains, with coloured pathways tracing charge transport between them.
Front cover, Phys. Chem. Chem. Phys. (2016) — “Planarity and multiple components promote organic photovoltaic efficiency by improving electronic transport.” Coupled classical and ab initio molecular dynamics showing that polymer twisting suppresses both light absorption and hole transport.
accurate rough compute cost → conventional attenuated MP2 same accuracy, orders of magnitude cheaper errors −80% · compute cost cut by orders of magnitude · 4M CPU-hours won competitively
The whole job is moving left without falling down: a cheap calculation that still tells the truth.

Credentials

Education
Ph.D., Computational & Theoretical Chemistry, UC Berkeley · B.S. Chemistry, Butler University
Patents
Multi-engine ASR (US 10,930,287) · Body-worn camera audio analysis (US 12,014,750) · Behavior detection (US 20240169854A1, pending)
Publications
18 peer-reviewed papers — Google Scholar. Journal front cover: “Planarity and multiple components promote organic photovoltaic efficiency by improving electronic transport”, Phys. Chem. Chem. Phys. (2016)
Open source
Contributor to vLLM and Hugging Face Transformers
Tools
Python, SQL, C++ · PyTorch, Transformers, vLLM, Whisper · Kubernetes, Postgres, GCP