Pat Pham / AI

TPM at Google / YouTube.

Technical Product and Program Manager building local-first AI tools and custom personal agents.

Selected Works

PatOnPoints↗

Live

An award travel search engine that routes between flight availability lookups and RAG-based booking advice.

  • 14 strategy guides in the source corpus
  • 2 query paths: flight search and strategy advice
  • • Combines flight availability, cash price benchmarking, and booking instructions in one flow
  • • Uses embeddings-based routing to show structured tables or RAG chat based on query intent
Next.js Express SQLite Duffel API +2 more
Deep Dive →
Live Site ↗ Private Source

Baymax Household Chief of Staff

In active household use

A persistent household operating layer built on Hermes Agent, combining routed models, layered memory, deterministic data pipelines, and human-approved action loops.

  • 12 enabled household and runtime jobs
  • • Acts as a persistent chief of staff across calendars, health, newborn routines, finances, household milestones, and developer workflows
  • • Protects prompt cache by disabling advanced tools on chat gateways and isolating concurrent classification
Hermes Agent Honcho LightRAG Langfuse +4 more
Deep Dive →
GitHub ↗

Atlas

Paused

A private place recommendation engine that aggregates Google Takeout history and group taste profiles into explainable suggestions.

  • 3 recommendation inputs: taste, friends, and places
  • • Aggregates friend location signals to beat anonymous public review averages
  • • Caches Google Places API data locally in SQLite to reduce lookup overhead
React Express SQLite OpenRouter +2 more
Deep Dive →
Private Source

Countergravity

Prototype in use

A mobile web interface and secure Tailscale tunnel for monitoring and steering local AI development environments.

  • 3 components: mobile UI, proxy, and private tunnel
  • • Bridges desktop coding workspaces to mobile screens for on-the-go review
  • • Secures local port connections via a private Tailscale tunnel
JavaScript Caddy Tailscale PWA +1 more
Deep Dive →
Private Source

PCloud

Tinkering & Benchmarking

A personal cloud of locally hosted LLMs and automated benchmarking pipeline on dedicated remote hardware.

  • 5 evaluation tasks per benchmark suite
  • 220W configured GPU power limit
  • • Automates remote llama-server starts and power caps on dedicated RTX 2080 Ti hardware
  • • Isolates model completions and chat endpoints to prevent benchmarking parser errors
Python llama.cpp lm-eval nvidia-smi +1 more
Deep Dive →
Private Source

Reusable Workflows

Skills built through daily Codex use

The most useful agent skills encode judgment: when to simplify, when to delegate, what to remember, and how to verify a release. These are six workflows I use and maintain across projects.

patppham/skillz-n-thingz

Agent Skills Library

Reusable prompts and workflows for planning, review, testing, and operations.

View Repository →

Ponytail

A family of workflows for reducing unnecessary code: choose the simplest implementation, review a diff for complexity, or rank what an entire repository can delete.

ponytailponytail-reviewponytail-audit

Selective orchestration

Delegate independent work with explicit file ownership and acceptance criteria. Keep the primary agent responsible for integration, verification, and the user conversation.

orchestrate

Repository memory

Turn verified commands, architecture boundaries, and project quirks into compact AGENTS.md guidance that helps the next coding session start accurately.

init

Release discipline

Reuse checks that already passed, document meaningful changes, and verify deployment after an explicitly requested release.

push