Skip to content
Kiran

About

I architect production AI systems end to end. Then I measure them honestly, even when that kills the brief.

Asked to build a model router, I measured the fleet first and found that model choice was about one percent of the bill. The other 62.9% was cache writes. The recommendation I delivered was smaller than the one I was asked for.

How I work

I measure the thing before I optimise it

Twice now, auditing my own safety net found it was measuring nothing at all. Once it was an evaluation suite whose assertions did not bind. Once it was a gold set recording the model's own predictions as ground truth. Both read as evidence. Finding them was worth more than any model I have shipped.

I would rather recommend less work

The most useful thing I have delivered was an argument for a smaller version of what I had been asked to build. It is not a comfortable position to hold in a review, and it is usually the correct one.

I write down what I disagreed with

Disagreements that live in the deliverable are reviewable. Disagreements that live in someone's memory get re-litigated with different facts six months later.

I attach the limit to the number

Every figure on this site carries how it is known. A measurement without its validity condition will be reused somewhere it means nothing, and by then it is not my sentence any more.

What I am good at

Agentic AI in production

Multi-agent systems that do not run away, do not leak, and do not lie.

Budgets at the tool-call boundary, wall-clock deadlines that survive a human-approval pause, durable checkpoints on evictable infrastructure, cross-instance run control, and authorization that fails static rather than open. Every one of those exists in my systems because something specific broke first.

Measurement and evaluation

Knowing whether an AI system works, and being allowed to say so.

Validity gates that derive what a report may and may not claim, paired statistics with an effect-size floor, judge-reliability sampling, a written retraction policy, and audits of the instruments themselves. Twice, auditing my own safety net found it was measuring nothing.

AI platforms and economics

Making a fleet of agents affordable and operable.

Model routing as an operational lever rather than a code change, cache economics measured rather than assumed, blast-radius floors, and offline pre-computation that makes an expensive question cheap before an agent ever asks it.

Machine learning

Models that beat the incumbent, measured honestly.

Hybrid ensembles with deliberately asymmetric detectors, hierarchical fallback when the precise comparison group runs out of data, grain-and-fanout gates between extraction and modelling, and baselines implemented in code so the improvement is a comparison rather than an assertion.

What I will not claim

A portfolio that only lists strengths is not evidence of anything. These are the honest edges of what I have done.

I have not managed people

No direct reports, no hiring, no performance management. I lead technically and across teams, and that is a different job. If you are hiring for the other one, I would be learning it on your time.

Most of what I have shipped is internal

Enterprise systems inside one organisation, behind SSO, serving colleagues rather than customers. I have not run a public-facing product at consumer scale, and the failure modes there are not the ones I have lived through.

Some of my strongest numbers are attested, not instrumented

The freight pricing figures were reported by the business rather than read off a dashboard I built. They are labelled that way everywhere on this site, and I would rather they read as weaker than they sound than have them quoted as measurements.

You cannot inspect any of it

Every system here is described under a public alias, with no internal names, endpoints, schemas or source. That is the right call and it is also a real limitation on what this site can prove. Ask me to whiteboard any of them and the detail will hold up.

What I keep doing

The seven patterns that show up across everything I build. Each one is argued for, and evidenced, on the architecture page.