Nate Bowers

Ask a general chatbot about Huberman's protocols and it will confidently invent half of them. Huberman GPT only answers from what he has actually said — 342 episodes, ~800 hours, 7.3M words — fusing semantic and keyword search so every claim cites the exact moment in the video.

• Next.js
• TypeScript
• RAG
• Supabase (pgvector)
• Hybrid Search
• Groq
• Vercel

• · 342 episodes · ~800 hours · 7.3M words
• · Hybrid retrieval (semantic + keyword) fused with RRF
• · Finds a relevant source ~99% of the time (LLM-judge eval, 100 questions)
• · Citations deep-link to the timestamp; runs on free tiers

Most AI newsletters are one scheduled blast written for an average reader. dailybriefmail builds one per person: you compose from 22 modules, and each morning a cron job turns your picks into live web-search instructions for Claude, which searches, returns structured JSON, and ships the brief through Resend.

• Next.js
• Supabase
• Claude API
• Resend
• Vercel Cron

• · 22 modules to compose from
• · A fresh brief per reader, every morning
• · Live at dailybriefmail.com

Surf forecasts are noisy, and the NOAA buoys meant to ground them report on their own irregular, gappy intervals while the Open-Meteo forecast feed runs on a regular grid. The real work is reconciling those mismatched timelines into one clean, comparable signal you can actually score a break on.

• Python
• Open-Meteo
• NOAA NDBC
• PostgreSQL
• Streamlit

• · Scores 3 NorCal breaks, 0–100
• · Score blends wave energy, wind, and swell angle
• · Best 2-hour window each day
• · Runs daily; tracks error (MAE + bias) vs. buoy