Gemini Decisions API: what it is, when to use it
Google's Gemini Decisions API scores yes/no, choice and rubric questions in one pass. What it costs, where it fits, and where a chat model still wins.
Blog
Notes from building apps, shows and AI systems with a small team.
RSSLatest
All posts
How to report a local problem in India so it gets fixed
A plain guide to reporting a civic problem in India: what to write, which office to use, the official deadlines, and how to follow up when nothing happens.
Agent Skills left Anthropic. Write yours like this
On 22 October Anthropic donated Agent Skills to the Linux Foundation's AAIF. What a skill actually is, why we keep ours short, and what not to install.
Independent evals after Anthropic let METR in
On 15 October Anthropic gave METR employee-like access to its models. What independence really means, and a version a small product team can run.
Why a chatbot that agrees with you can lead you wrong
Two studies found AI chatbots flatter users and do not help people decide better. How to ask an AI for advice without being told what you want to hear.
Claude Haiku 5.5: when the cheap model is enough
Haiku 5.5 costs about 75% less than Haiku 4.5. What we would move onto it in Cabin, what we would not, and why evals come first.
How to stop overthinking a decision: a 6-step method
How to stop overthinking a decision: name the question, set a time limit, define good enough, cut the options, get distance and act. With sources.
Why Anthropic asked a Vedanta monk about Claude
Anthropic met religious scholars, including Swami Sarvapriyananda, about Claude's values. What we know, what it means, and lessons for AI teams.
5 Indian apps to try in 2026: decisions, quizzes, jobs
Five Indian apps from one small team: Cabin for decisions, QuesMo for friend quizzes, Karma, an interview prep app in beta and a local shop finder.
Running a 24/7 AI software factory: builders and one advisor
How we set up a team of AI coding agents and one advisor agent to ship tested tickets around the clock: roles, ledgers, release trains and what broke.
Social quiz apps: why questions beat passive scrolling
What a social quiz app changes: answering with friends instead of scrolling alone, and what research on active and passive social media use says.
How we ship software with AI agents without losing control
Specs, small plans, task-scoped agents, independent review and real checks. How agents do the work while people keep the decisions.
How to design a responsible AI companion app
Practical rules for a responsible AI companion app: say it is an AI, skip engagement hooks, handle crisis moments, and give people control of memory.
Building a Hinglish AI chatbot: lessons for India
How to build a Hinglish AI chatbot people enjoy: match the user's register, handle Roman and Devanagari script, test code-mixed prompts and watch cost.
Routing between LLM providers to balance cost and quality
Why we route requests across more than one model provider, how prompt caching cuts repeat costs, and why routing changes pass offline evals first.
llms.txt in 2026: what it does and what it doesn't
A practical llms.txt guide: what the file is, what Google says about it, who uses it, and how to generate one from your site's content in minutes.
Designing AI memory that stays accurate
Long-term memory for an AI companion is about keeping facts current, not storing more. Recall, supersession, and why contradictions cost trust.
AI content labelling rules in India: a creator's guide
India's AI content labelling rules and YouTube's synthetic content label, explained for creators: what counts, what to disclose and a simple checklist.
LLM evals for small teams: a practical starting point
How to build LLM evals with a small team: start from real failures, use pass/fail checks, calibrate an LLM judge and gate every prompt change.