I build machine learning systems that run in production and quantum algorithms that report what they actually achieved. Most of the work below comes down to the same thing: making a system's real behaviour visible, whether that means fixing a training run that dies silently or finding out that a 97% accuracy figure was measured on the wrong distribution.
Every engagement ends with a written account of what was done, what it measured, and what it did not cover.
Each with the work it comes from
Multi-GPU training that hangs, corrupts checkpoints, or silently produces nothing. Cloud training setup on Vertex AI and comparable platforms. Launcher topology, DDP configuration, tokenisation races, checkpoint integrity, and the monitoring that catches a dead run before it burns a weekend of compute.
Built the distributed Longformer training stack on GCP Vertex AI after a silent deadlock cost twenty hours of GPU time. See the work
An independent read on whether a reported number means what it appears to mean. Data leakage, checkpoint-selection bias, train/serve skew, benchmarks that measure something adjacent to the deployment case. Delivered as a written report with the reproduction steps.
Found and published data leakage in my own result, withdrawing a +1.4 point gain that survived re-evaluation at +0.1. See the work
Comparisons built so that a negative result tells you which component is at fault. Controls, baselines that can fail, significance testing across splits, and honest reporting of what the measurement does not cover.
Built an exact brute-force control that proved a failing quantum solver was a formulation problem, not an optimiser problem. See the work
Hybrid quantum-classical pipelines in Qiskit. QUBO formulation, variational solvers, Grover oracles, encoding schemes, noise-robustness studies, and resource analysis that reports measured circuit depth rather than asymptotic promises.
Qiskit Advocate at IBM. Reproduced a reference RNA structure at zero energy gap across 35 noise conditions. See the work
Making a technical result legible to people who did not build it — research write-ups, documentation, interactive explainers. Written so a reader can check the claims rather than take them on trust.
Wrote the interactive explainer and final report for the WISER Moderna challenge. See the work
Hands-on sessions on quantum computing and machine learning for student groups, developer communities, and teams starting out in either. Materials built to be reused after the session ends.
Ran a TensorFlow workshop for Google Developer Group on campus. Technical Team Lead at QQuEST.
First
Tell me what the problem is and what you have already tried. If it is not something I can do well, I will say so and point you elsewhere.
Then
A scoped proposal with a fixed deliverable, a timeline, and a price. For audits and short engagements this is usually a fixed fee rather than an hourly rate.
Throughout
Working code and written findings as they land, not a single delivery at the end. If something I find changes the scope, you hear it when I find it.
Send a short description of the problem. I reply to everything, including to say no.
Currently taking contract and collaboration work · Pune, India · remote