AI BUILDS: RAG, AGENTS, AUTOMATIONS

My mom's health insurance claim was rejected on a vague excuse with no real explanation. I built ClaimRight to fight it, used it on her case, and got the claim reopened in two days.

• whatRAG and agent systems I build on my own, and ClaimRight, the biggest of them.
• my partAll of it. Retrieval pipeline, thresholds, the regression suite, and the case-management app around them.
• result10-plus tools shipped with 500-plus users between them, and my mom's claim reopened in two days.

Why I built ClaimRight

Her rejection letter was the kind written to make you give up: a clause reference, no reasoning, no route to argue with it. I pulled the IRDAI circulars and the ombudsman precedent that applied to her policy, wrote the appeal off them, and the claim was reopened within two days. The insurance manager called and apologised personally. The money is still not back. It is at the final ombudsman stage waiting on a date, on a process that runs about a year.

The market's numbers, which are not mine, are why the tool is worth building at all:

• ₹26,037 crore of health insurance claims rejected or disallowed in India in FY23-24.
• Under 1% of rejected claimants ever fight back.
• 87% of the ones who do, win.
• 94.5% of cases that reach the insurance ombudsman get resolved there.

Most people never appeal because they have no idea which regulation their insurer just broke.

How it works

ClaimRight reads a denial letter, retrieves against a knowledge base covering nine IRDAI claim categories held in Supabase pgvector, and drafts an appeal where every line is tied back to a source document: the circular, the policy wording, or the ombudsman award it rests on. The gates are the part I would defend:

• A 0.65 similarity floor, under which a passage is not retrieved at all.
• A 0.40 citation-cut threshold, under which a drafted line loses its citation and gets dropped instead of shipped uncited.
• A 27-case regression suite the thresholds are tuned against. Each of those numbers is one I tested into place.

I built it in Claude Code. Early on, a senior systems engineer walked me through the architecture, which is most of why the retrieval layer is structured the way it is.

The other builds

ClaimRight is the biggest one. The rest of the list:

• Hultivate CMF, a working custom GPT that walks someone through a full career plan. I used it on my own career and built one for a friend.
• The internal and outbound AI agents at MTech-X that ran the sales and management workflows.
• The canvassing intelligence layer on the Congress campaign that read call logs and flagged failing scripts.
• Data-visualisation tools, n8n automations, and custom Claude skills, built one at a time for whatever problem was in front of me.

Ten-plus tools, 500-plus people using them. Most of them, plus the community work around them, sit in one place.