I have been the person on call for this.

I have spent seven years as an ML architect and technical lead at Amazon. The systems I owned served two million requests an hour at 99.99% availability. On those systems I cut inference latency by 80%, built a classifier over 100,000+ classes that raised accuracy by 35% and halved processing time, and delivered more than $20M in savings.

Before the ML years, I spent nine years in Amazon operations, programme management, and product. That matters more than it sounds. It means I design for the person doing the work and for the cost of running the system, not just for the benchmark.

I have sat on both sides. I have been the person who has to make a system work at scale, and the person who has to explain to a business why it is not working yet. I know what vendors leave out because I have had to fix what they left out.

You get me, not a delivery team. The person who scopes the diagnostic is the person who does the work, and the person you can reach while it is running.

Why independent

I do not sell software. I do not take referral fees from vendors. I do not have a partner programme. Where the right answer is a product you buy without me, I say so — in writing, in a document you keep, so you can hold me to it later.

A recommendation from someone who profits either way is worth less than the paper it is printed on. That is the entire reason this practice is structured the way it is.

How I work

1. Diagnose the whole stack. Data and retrieval, model and inference, orchestration and evals. Most production failures are not prompt problems.
2. Fix the highest-leverage problem first. One change that moves cost, latency, or reliability, measured against a baseline agreed before the work starts.
3. Ship with evals, guardrails, and observability. A system that fails loudly, names its failures, and can be operated by your team without me.
4. Hand off with runbooks. Written architecture decisions, cost model, failure modes, and on-call notes. Your team keeps the code.

What I do not do

• Hourly billing, staff augmentation, or body shopping.
• Free discovery calls, free proposals, or free proofs of concept.
• Generic AI strategy with no access to systems or data.
• Publish client results or testimonials before a client has approved them.

Location and availability

I am based in Chennai, India, with an 8–11am ET overlap. I have worked with US teams for 16 years. If daily onsite presence is required, I am not the right fit — and I will say so before you buy anything.

I keep the practice small — a handful of engagements at a time — so the person who scopes the work is the person who does it. If the timing does not fit, I will tell you rather than take the work and subcontract it.

References from Amazon colleagues are available on request, labelled as professional references rather than client testimonials. I sign NDAs, I do not train on client data, and I work inside your environment.