Writing

A voice agent is not a chatbot that happens to speak. It is at least six coupled systems running under a latency budget measured in milliseconds, and most of what makes an agent feel robotic traces back to how those systems are stitched together.

How a college RAG experiment became a job-matching and skill-gap engine integrated with India's National Career Service: the matching approach, the skill-gap modelling, the data constraints, and what I would redesign.

The most useful thing an extraction step can output is nothing. A note from building matching and lead-scoring pipelines on messy data.

What it actually costs to deliver AI automation at ₹1L+/month revenue: model spend, delivery time, and the maintenance tail nobody prices in.

What actually breaks when you put agents into WhatsApp-run, person-dependent business processes, categorised from real deployments.

Automating a business that runs on WhatsApp and one person's memory. The model was never the hard part.

Anonymised, aggregate patterns in where students are actually short of job readiness, drawn from the university pilot cohorts.

A skill-gap is not a list of things you don't know. It's the ordered set of things that would change an outcome, and the ordering is the whole problem.