Gooey AI Yield
A value-measurement and coaching platform for AI-assisted engineering — quantifying the real ROI of AI coding tools rather than just counting tokens.
Highlights
- Designed a Value-Per-Token metric and a 0–100 AI Efficiency Score that diagnoses rather than stack-ranks engineers
- Built a metadata-first, privacy-by-design telemetry pipeline across five capture surfaces, including a Claude Code plugin
- Implemented AES-256-GCM column-level encryption with searchable blind indexes for all PII
Skills
Overview#
Gooey AI Yield is a platform I designed and built to answer a question most teams adopting AI coding tools can't yet answer: is the AI actually creating value, or just generating tokens? Instead of vanity metrics, it measures the value of AI-assisted work through a Value-Per-Token (VPT) metric and a 0–100 AI Efficiency Score, coaches engineers in-tool, and gives managers an ROI-first dashboard.
The guiding design principle is that the score diagnoses, it does not stack-rank — and that data capture is metadata-first and privacy-by-design from the ground up.
The Problem#
Organizations are spending heavily on AI coding assistants with little visibility into whether that spend translates into durable, shipped value. Raw usage stats reward activity, not outcomes, and naive measurement creates surveillance fears that erode trust. The challenge was to build something that measures real value, survives engineer scrutiny, and never becomes a productivity-policing tool.
My Role#
Founder and sole engineer. I own the product design, metric model, backend, web app, mobile companion, capture surfaces, and cloud infrastructure end to end.
Architecture & Approach#
A single repository houses five independently-deployable surfaces sharing one environment configuration:
- API — Fastify + TypeScript on PostgreSQL, with a domain model spanning organizations, teams, capture events, surviving outputs, and the derived VPT and efficiency-score engines. Scores are computed from captured telemetry plus "surviving output" (merged PRs / closed tickets tracked over a survival window) with cohort normalization across refactor, greenfield, and bugfix work.
- Web — a React 19 + Vite SPA with a light marketing site and a dark product app behind invite-only auth, built on a custom design system.
- Mobile — a read-mostly Flutter companion at rough parity with the engineer self-view and manager glance.
- Capture — five real telemetry surfaces (a Claude Code plugin, a VS Code extension, an MV3 browser extension, an Electron desktop agent, and a CLI wrapper), each posting metadata-first events through an authenticated ingestion gateway.
- Infrastructure — a Terraform Cloud–managed AWS stack (ECS Fargate, RDS, S3 + CloudFront, Secrets Manager, SES).
Technical Highlights#
- Privacy-by-design ingestion. The events gateway authenticates each surface with a per-organization HMAC capture token and rejects content fields unless the user has explicitly opted into content capture — privacy is enforced at the validation layer, not as policy.
- Column-level encryption. PII (emails, names, organization names) is encrypted with AES-256-GCM using a versioned envelope and per-column additional authenticated data. Searchable fields get an HMAC-SHA256 blind index so encrypted data stays queryable without exposing plaintext.
- Derived-metric engine. The VPT and scoring engines transform raw capture events and surviving outputs into four coachable subscores, with per-organization configurable weights, windows, and thresholds.
- Invite-only auth with email/password, OAuth, and magic-link flows and no public registration endpoint.
Skills Demonstrated#
AI/LLM tooling and instrumentation, original product and metric design, secure backend architecture, applied cryptography for data privacy, and full-stack delivery spanning web, mobile, and infrastructure-as-code.