Writing
Notes from the build.
Working notes on the parts of AI engineering that are genuinely hard: retrieval that fails quietly, evaluation that stops being valid, voice pipelines that meet real Indian speech, and what Indian regulation actually asks of a system that processes personal data.
These are engineering notes, not thought leadership. Each one answers a question in its first two sentences and then spends the rest of the piece justifying the answer, including the parts that argue against us.
Our earlier writing covered blockchain topics and has been retired — it no longer reflects what we do, and leaving it up would have been a claim about our current practice that is not true. What replaced it is narrower: problems we have hit repeatedly on client work, written up with the diagnostics rather than the conclusions alone.
Where a claim is regulatory or measured, the source is named in the text and linked at the foot of the article. Where we do not have a number, the sentence is written without one. We would rather publish a handful of pieces a year that are worth citing than a weekly post restating vendor documentation.
Articles
Everything published so far
What the DPDP Act actually requires of an AI system
The obligations that survive contact with a vector store, a prompt log and a fine-tuned model — and the one that quietly makes fine-tuning on personal data a bad idea.
RBI data residency and calling an external LLM
Payment data, the 24-hour rule, the outsourcing direction’s right-to-audit clause, and why redaction at the boundary is a design problem rather than a regex.
Why voice agents fail on Hindi-English code-switching
Where the transcript actually breaks, why word error rate is the wrong target, and what has to change in endpointing, normalisation and evaluation.
Why your RAG system gives wrong answers, in order of likelihood
Eight causes, ranked, each with the cheapest test that distinguishes it from the one above — ending, reluctantly, at model choice.
How to evaluate a RAG system before you trust it in production
Every metric worth reporting, with the failure it does not catch, a golden-set schema you can copy, and a retrieval gate that returns a non-zero exit code.
The gap between an AI demo and something that runs unattended
Retries that send the email twice, agents with more authority than anyone intended, and the alerting signals that actually catch a failing agent.
How to choose an AI development company, and when not to hire one
Acceptance criteria that can actually fail, who owns the prompts and the evaluation set, the disqualifying answers, and the checks we would fail ourselves.
What SEBI’s rules actually require of a stock broker’s technology function
Reporting windows a human process cannot meet, capacity stated as a multiple of the peak, retention on three separate clocks, and the accountability that does not move with the work.
The service pages carry the same material in a different shape: how we approach retrieval systems, voice agents and agents that take actions.
Argue with any of this
If one of these pieces is wrong about your situation, that is a useful conversation. Bring the counter-example and the specifics of your system.