The Pareto line: how to pick an AI model without overpaying
Most people read AI model charts top-down and pick the highest dot. The line along the top-left edge is the part that matters. A plain-English guide to the Intelligence Index, cost per task, and using the Pareto line to pick a model at any budget.
Oct 13, 2026Yeh Honsla: A Tribute to Salim-Sulaiman
My cover of Yeh Honsla from Dor, and a short film about an astronaut finding his way home. A tribute to Salim-Sulaiman, whose music stands beside you when things fall apart.
Oct 11, 2026Banjaara: A Tribute to Mithoon
Mithoon composed it, wrote it and sang it. Why Banjaara from Ek Villain still feels like someone quietly took the hurt away.
Oct 10, 2026EmbeddingGemma 2 puts multimodal search on the device
Google's EmbeddingGemma 2 puts text, photos, audio and video into one vector space and runs in about 567MB on a phone. Five use cases worth building, the numbers, and the limits to test first.
Oct 10, 2026Haiku 5.5 thinks like a flagship and bills like a budget model
Claude Haiku 5.5 lands in the same intelligence tier as Kimi K3 and Opus 4.8 on Artificial Analysis, at roughly a ninth and an eighteenth of the cost per task. What that changes for model routing, and what the benchmarks don't tell you yet.
Oct 8, 2026Adding agents to software you already run
A field guide for teams adding agentic capabilities to an existing platform with open models: when to pick LangGraph, Agno, Pydantic AI or Mastra, where CrewAI, Langflow and Dify fit, and what to check before you ship.
Oct 7, 2026Maaeri: A Love Letter to the Indipop Years
A piano and vocal cover of Euphoria's Maaeri, and a small thank you to the era that made Hindi pop feel like it belonged to us.
Oct 2, 2026Ae Dil Hai Mushkil: The Strength of Loving One-Sided
A cover of the Ae Dil Hai Mushkil title track, and why one line from the film has stayed with me for ten years.
Oct 1, 2026Every lab's models misbehave. OpenAI's keep getting out.
Frontier models from every lab broke the rules under test this summer. The difference is who caught it: for OpenAI, it was usually the people on the receiving end.
Sep 29, 2026Dil Ibadat: borrowing from KK again
A piano and a microphone, played and sung live, and a song that was never mine to keep. Notes on stripping a KK anthem back to its bones.
Sep 26, 2026When the regulator asks what your agent did, your prompt isn't the answer
Spain's data protection authority has logged a breach it says was run end to end by an AI agent. The details are thin. The lesson for anyone shipping agents isn't.
Sep 17, 2026The coding harness argument is really an argument about defaults
Same model, same benchmark — Claude Code scores 42, opencode scores 51. The nine-point gap is real. It was also bought with 65% more tokens, and Claude Code's own effort dial closes it. The harness isn't the variable people think it is.
Aug 26, 2026The cheapest AI upgrade we found was a settings change
A cost-optimized model at maximum effort scores 52 on the Intelligence Index for five cents a task. The flagship scores 61 for ninety-six. The interesting part isn't the gap — it's what you do with it.
Aug 25, 2026What I learned restructuring an engineering org around AI agents
We didn't add agents to our workflow — we rebuilt the workflow around them. Here's what actually changed, what broke, and what I'd do differently.
Jul 5, 2026Every song's a frame: how I direct AI music videos
The Kalank cover started as a vocal take and became an animated film — a wanderer, a spirit of golden light, an impossible bridge. Here's my scene-by-scene process.
Jun 20, 2026Mixing consoles and microservices — one discipline, two crafts
Twenty years of doing both taught me they're the same job: take a system of many small decisions and make it feel like one inevitable thing.
May 28, 2026