Now open for select engagements — Zach is a free agent

ZACH FARR

AI product engineer & systems builder

zsh

zach@studio ~ % echo "Hello, world."
I build AI-powered products, rich frontends, and reliable backends.
AI product engineer & systems builder.

type `/help` to list projects

Preview

How it started – 2013

I learned by taking things apart. View‑source, copy, break, fix. That bias for reverse‑engineering became a habit of shipping: short loops, fast feedback, and relentless polish.

Early experiments: canvas graphics, CSS grids, scraping
First products: small tools for friends that solved real pain

Shift to systems: testing, deploy pipelines, observability
Today: AI‑assisted loops from zero → users → iteration

Selected Projects

End‑to‑end work across product, AI, and systems.

zachcode

Sales Call Analyzer

for Apex Acquisition

End‑to‑end system that turns raw audio into structured insight: speaker diarization, sentiment trends, and action extraction with a tight UX. Built for fast operator workflows and low‑friction review.

React
Next.js
TypeScript
Python
FastAPI
Docker
Npm
Vercel
Railway
OpenAI
Deepgram
Supabase

Highlights

Streaming transcription → segment alignment → insight generation in one pass
Deterministic action extraction prompts with evaluation harness
Zero‑copy data flow, signed URLs, least‑privilege keys
Polished UI patterns: keyboard nav, instant filters, deep‑links

−63%

Avg. review time

<400ms

Cold start

1.9s

Latency (p95)

99.9%

Uptime

Story – after Sales Call Analyzer

1Who was the primary operator and what did they need to do faster?
2What made raw calls hard to review before this?
3Which constraints drove the architecture (latency, cost, privacy)?
4Why FastAPI + Deepgram + OpenAI + Supabase, and what did I reject?
5Where did v1 break and how did I fix or simplify it?
6What metric proved it worked (time‑to‑review, accuracy, retention)?
7What’s the next improvement I’m excited to ship?

zachcode

PrimeCop

2016

Chrome extension that automated checkout with resilient DOM strategies and safe defaults. Treated as a product: fail‑closed logic, explicit states, and responsive UI.

JavaScript
Chrome
jQuery
Bootstrap

Highlights

Background script orchestration with message‑passing workers
DOM diffing to survive layout shifts and anti‑bot changes
Configurable rules with human‑in‑the‑loop confirmations
Telemetry to tune timings and reduce false negatives

3k+

Users

Weekly

Release cadence

<0.2%

Crash rate

<3s

Core flow

Story – after PrimeCop

1What pain did PrimeCop remove on drop day?
2How did I keep automation resilient to DOM and anti‑bot changes?
3What safeguards protected users (fail‑closed, confirmations, delays)?
4How did I distribute, support, and update for thousands of users?
5What did I learn about platform rules and automation ethics?
6One production bug that taught a lasting lesson.

zachcode

Keepr AI

demo in progress

Bookkeeping copilot that reconciles transactions, drafts categorizations, and answers natural‑language questions over ledgers. Designed for traceability and auditability first.

React
Express
OpenAI

Highlights

Schema‑grounded prompts that emit typed events and diffs
Human reconcile queue with confidence thresholds and reasons
Ingestion → normalization → vector‑backed search
Explain‑why answers with links to source rows

−50%

Manual effort

−35%

Time to close

<0.5%

Hallucinations

100%

Coverage

Story – after Keepr AI

1Which bookkeeping workflows are in scope for v1?
2How do I ground model output in an accounting schema?
3Where do humans stay in the loop and why?
4How do I measure correctness and prevent hallucinations?
5Which data sources are required and how is privacy handled?
6What makes the UX trustworthy for finance work?

zachcode

WagerMind

Expo / React Native

Sports picks product with an opinionated UX for confidence and bankroll management. Shared services and model orchestration with a React Native client.

React
Next.js
TypeScript
Python
FastAPI
Docker
Npm
Vercel
Railway
OpenAI
Deepgram
Supabase

Highlights

Model ensemble with guardrails and calibration curves
Feature store for rolling stats; scheduled refresh jobs
Bankroll policies and risk controls encoded as reusable rules
Shareable insight cards with compact data viz

+22%

Retention (8w)

<100ms

FE perf (p95)

~250ms

API latency

99.8%

Crash‑free

Story – after WagerMind

1Who is the target bettor and what job are they hiring this for?
2How do models combine and get calibrated?
3What bankroll policies and risk controls are enforced?
4Which signals proved sticky for retention?
5Trade‑offs between transparency and speed on mobile?
6What’s the next experiment on the roadmap?

Operator’s manual

Principles I build with when the stakes are high and the loop needs to be fast.

Ship a usable core before adding breadth. Small, lovable, and complete beats large and unfinished.
Make truth observable. Logs, traces, metrics, and product analytics wired in from day one.
Bias to clarity. Code and UI that explain themselves; explicit over implicit; sharp edges documented.
Model‑aware UX. When AI is involved, design for confidence, reversibility, and feedback.

Contact

Interviewing for roles in AI product engineering and full‑stack systems. If you’re hiring, I’d love to connect.

Email for roles