Harshil Jani·4d agoHaversine vs OSRM: How Far Apart Are Two Places? (A Bug I Shipped)“Her pickup point is 6.2 kms away from her address, and she wants to cancel.”
Harshil Jani·May 21Shipping LLMs (Part 6/6): How to Stop an LLM Agent From LoopingPreviously: Shipping LLMs (Part 5/6): Where Your LLM Tokens Actually Go. I named the five silent token leaks. This piece is about the…
Harshil Jani·May 20Shipping LLMs (Part 5/6): Where Your LLM Tokens Actually GoPreviously: Shipping LLMs (Part 4/6): How to Evaluate a RAG Pipeline. I argued for the 90/10 RAGAS-plus-human eval rhythm. This piece is…
Harshil Jani·May 17Shipping LLMs (Part 4/6): How to Evaluate a RAG PipelinePreviously: Shipping LLMs (Part 3/6): Speculative Decoding vs Quantization. I argued you should run both. This piece is about whether the…A response icon1A response icon1
Harshil Jani·May 17Shipping LLMs (Part 3/6): Speculative Decoding vs QuantizationQuantization fixes memory bandwidth. Speculative decoding fixes autoregression. Stack them for 3–4x cheaper LLM inference, in this order.
Harshil Jani·May 16Shipping LLMs (Part 2/6): What’s Actually in Your KV Cache?Previously: Shipping LLMs (Part 1/6): Prompt Caching vs Semantic Caching. I argued you should always prompt-cache your stable prefixes…
Harshil Jani·May 16Shipping LLMs (Part 1/6): Prompt Caching vs Semantic CachingA user types a question into your AI-powered support chatbot:
Harshil Jani·Apr 10Why Talking to Agents Is the Highest-Leverage Move in 2026 — By Anthropic’s Claude Code TeamI’ve been following Thariq Shihipar (@trq212) for a while now. He’s on the Claude Code team at Anthropic. He’s the guy who built the…
InArtificial Intelligence in Plain EnglishbyHarshil Jani·Mar 13How to Get Clean Text in AI-Generated Images Every TimeEvery LLM which I use Claude, ChatGPT, Grok etc. has the same infuriating problem when it comes to texts on images. I ask for a clean…A response icon1A response icon1
InProduction Engineering PlaybookbyHarshil Jani·Mar 6Stop falling in love with your projects and Ship them out.I’ve killed more side projects than I’ve shipped. Beautiful READMEs, clean architectures, clever abstractions and zero users. The graveyard…
