eric crook.

Eric Crook · San Francisco Bay Area

I build AI that shows up for work.

Agentic systems for the jobs nobody should still be doing by hand — voice agents, back-office automation, document pipelines — and the operations consoles that keep humans in control. 24 years of production engineering behind every line.

See the work Read the resume

Open to founding & staff+ roles · SF Bay Area / remote

24 yrsshipping production
1B+events / day
2.5Mdaily active users
−90%compute spend
$30M+raised alongside
2ml patents

Work · Product concepts

The shape of AI B2B, drawn by hand

Most of the last three YC batches converge on the same dozen product surfaces: agents that do a job, and consoles that let humans trust them. These are working concepts for those surfaces — every frame below is live markup, not a screenshot. Two are already full working demos you can drive; the rest are queued.

AI voice agents concept

Voice operations desk

A call floor for voice agents: live transcript synced to audio, sentiment timeline, outcome tags, and one-click barge-in when a human should take the call.

Human-in-the-loop live demo

Greenlight — approval queue

Agents propose, people dispose: every action lands here with its evidence and a calibrated confidence score. High confidence flows; low confidence waits.

open the live demo →

Agent infrastructure live demo

Stepwise — trace inspector

The debugger agents deserve: every LLM call and tool call as an expandable span, with latency, tokens, and cost per step. Failures cluster themselves.

open the live demo →

Vertical AI · claims & docs concept

Extraction review

Source document on the left, extracted fields on the right, confidence on every value. Reviewers only ever touch what the model wasn't sure about.

LLM quality gates concept

Eval scoreboard

Deterministic and LLM-as-judge suites trending over time — and wired into CI, so a regression in accuracy, latency, or cost blocks the deploy that caused it.

GTM automation concept

Outbound pipeline

An AI SDR that researches, drafts, and follows up — with every message passing a human review queue before anything carries your name outward.

RAG over private data concept

Knowledge desk

Natural-language answers over a company's own data, where every claim carries a citation that resolves to the exact source passage. Trust is a feature.

LLM ops & finops concept

Spend governor

Token and dollar attribution per task, user, and tenant — with budgets, caching, and alerts. I've cut an inference bill 90%; this is how you keep it cut.

Experience

Twenty-four years of receipts

Founding-engineer speed with CTO-level scope: games at millions of DAU, analytics at a billion events a day, and lately, agents that earn their keep.

Flowtel YC W25Founding Engineer
Low-latency voice booking agents for hospitality: live rates, card capture through secure vaults, multi-persona simulated-customer evals. Onboarding to live agent in under 5 minutes.
2026
math.technologyFounder
End-to-end LLM analytics stack (Python + Rust + Polars) over 2 GB/day of events; ~90% cheaper than the pipeline it replaced. RAG over customer data via Slack; vLLM on EKS; dSPy + eval gates.
2025 —
MoescapeStaff Software Engineer
Search and trust-&-safety over Stable Diffusion images and multi-LLM chat logs; shipped image-categorization models; saved thousands/month on BigQuery.
2024
MidnightPrincipal Software Engineer
Multi-chain digital-goods logic — mint, trade, destroy, forge — on-chain and in-game, with security first.
2022 – 23
DorianChief Technology Officer
Led engineering for a no-code interactive-fiction platform. Built the analytics stack, ran AWS infra and GDPR, supported the $14M Series A.
2019 – 23
Megastar MillionaireDirector of Engineering · ASX:MSM
Head of engineering and analytics at a publicly traded startup; ML models for categorization and prediction. Two machine-learning patents.
2016 – 19
Nexon Director → Principal · acq. Gloops
Built the core analytics system at 1B+ events/day; platform and data teams across two offices; backend for DomiNations at 2M+ DAU.
2012 – 15
Digital ChocolateSenior Software Engineer
Backend on Millionaire City at 2.5M DAU; lead backend across multiple 300K+ DAU games.
2008 – 12

The one-page version, ready to print: the resume · PDF

Stack

Tools that have shipped

Nothing aspirational below — everything here has carried production traffic.

Agentic / LLM

Claude CodeLangChainvLLMdSPy ChromaDBPineconeBedrockLLM-as-judge evalsLoRA

Languages

PythonRustGoTypeScriptJavaC#

Infra & cloud

AWS · EKS → KMSGCP + BigQueryTerraformDocker GitHub ActionsGrafana / LokiPrivateLink

Data

PolarsPyTorchPostgresRedshift KafkaKinesisAthenaMongoDB

After hours

The same engineering, pointed at fun

This domain used to be a wizard's tower. It still is — one floor down. Shipped games, playable now, built with the same care as everything above.

Card battler · live

Aether Vanguard

Three lanes, hidden reveals, a roguelike campaign — Phaser on an AWS backend.

play →

Virtual tabletop · live

Grimlight

D&D 5.5e for huge maps: server-authoritative fog of war, true multiplayer.

enter →

The old site · preserved

The Red Wizard's tower

The site this one replaced — embers, secrets, a wand for a cursor. Mind the shadows.

descend →

Contact

Building agentic products?

I'm at my best as a founding or staff+ engineer where the AI has to actually work — and pay for itself.

erictcrook@gmail.com LinkedIn