Build AI agents and orchestration systems, including the RAG tool that got my mom's rejected claim reopened in 2 days.

Most problems worth solving are a workflow underneath. Once you can write the workflow down, step by step, with the judgement calls named, it is usually straightforward to build something that runs it. That is the skill I have spent the last two years sharpening, and it is why the list of things I have built is as scattered as it is: a data-visualisation tool because I was tired of rebuilding the same charts, agents that ran outbound sales, a layer that read canvassing call logs. Whatever was in front of me.

Then my mom's claim was rejected, in a letter written to make her 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. ClaimRight is that process, built into a tool.

Compliance and regulation are the worst possible place to be casually wrong, which is most of why this one was hard. I use AI constantly and I still cannot stand slop. I have watched these systems hallucinate confidently enough to be believed, and making sure that does not happen is the thing I care about most in anything I ship.

For my mom's rejection I built the tool on retrieval-augmented generation, as a hybrid RAG weighted 70/30 between semantic similarity and text ranking. Pure vector search finds passages that read similar and will happily cite the wrong rule; legal text has terms that have to match literally, a clause number or a 36-month waiting period. Weighting both together is what keeps it honest.

Under it is a knowledge base of the regulation that actually governs these disputes, IRDAI circulars, policy wordings and ombudsman awards, chunked and embedded as vectors with the section and circular number carried on every chunk. That is the part that holds hallucination down and keeps the output inside the regulatory framework. A user uploads their rejection letter, the system retrieves against it, and the appeal comes back with each line tied to the source it rests on.

The rest of the bots get built the same way: write the workflow down, find the sources that are actually authoritative for it, and wire the thing so it shows its references instead of asking to be trusted.

• 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.

My mom's claim was reopened in two days. Beyond that, ten-plus tools shipped and 500-plus people using them.

The MTech-X agents ran a real sales workflow and handled internal knowledge compilation for the senior leadership. The canvassing layer moved a real team's persuasion rate from 35% to 50%. The rest exist because I hit a problem and building the tool was faster than working around it.

Most of this work is not the model. It is documenting the workflow well enough to automate it, then being honest about which steps have to stay human. Knowing where a system will hallucinate, and putting the guardrail at exactly that point, is most of the job.

The rest of it is the part that does not automate. Whether the output is any good, whether a person can take it somewhere and be taken seriously, whether it was worth building at all. Trust, taste and judgement are still yours to bring.