Run Claude Code with Your ChatGPT Subscription
Use your existing ChatGPT Plus/Pro/Max quota to power Claude Code via LiteLLM - no separate API key needed. Now with background launchd service and silent startup.
Prabal
AI engineer, consultant, and founder. Writing about LLMs, distributed systems, and building things that work.
Latest posts
Google Made Long Context 5x Cheaper
Google's TurboQuant compresses the KV cache by ~5x with minimal quality loss - what this means for open-source LLM inference.
LatentScore: Text to Real-time Music on CPU
An open-source Python library that turns text prompts into procedural music in real time, on CPU, sub-second.
The X Algorithm, Visualized
An RPG-style simulation of X's recommendation algorithm - pick a persona, write a tweet, and watch the pipeline decide if it goes viral. Real ML running in the browser.
Shipping in Production
Making LLMs produce reliable JSON - Part 3 of 3. Temperature tuning, graceful degradation, observability, token caching, jailbreak detection.
When One LLM Isn't Enough
Making LLMs produce reliable JSON - Part 2 of 3. Multi-model architectures, writer-critic loops, best-of-N sampling, and the freeform thinker + JSON formatter pattern.
Schema Design
Making LLMs produce reliable JSON - Part 1 of 3. Simple schemas, detailed descriptions, constrained decoding, and why enums beat strings.
The Post-Mortem
The post-mortem: 5 things that killed a bootstrapped B2B startup, what's salvageable, and what I'd do differently.
They Wouldn't Stay
Selling MonitorIntent to startup founders, the Clay defection, the vitamin problem, and watching our moat disappear in real time.
Scraping LinkedIn
Building MonitorIntent's MVP - 2 months to a Replit dashboard, 5x LLM cost reduction, and the realization that LinkedIn likes are not buying intent.
The AI Trust Deficit
The first product at MonitorIntent was clever - maybe too clever. AI messages that sounded human but made judgment errors nobody forgave.