Hardik Tankaria

I build the decision layer of AI systems: the part that works out what is worth computing, measuring, or asking next, and then proves it was worth it.

That question looks different depending on where you stand. In a self-driving laboratory it is which experiment should the instrument run next, and can we afford it. In computer-aided chemistry it is which of ten thousand candidate reactions deserves the next expansion. In a training loop it is how much curvature information is worth paying for. They are the same problem, and I work on all three.

What I am building now

I am an ML research engineer at MI-6 Co., Ltd. in Tokyo, where I own three concurrent R&D lines.

Autonomous experimentation. A closed-loop Bayesian optimization system for self-driving labs, built so the budget is structural rather than an afterthought: every proposed experiment is priced by a fitted cost model, charged against a ledger, and refused if the remaining budget cannot cover it. Twelve proposal policies, three interchangeable acquisition optimizers, architecture decision records, and a one-command reproducibility check that re-derives every committed figure from committed data. CNN encoders turn instrument images into surrogate inputs, so the loop learns from what the instrument actually sees rather than from a scalar proxy.

Agentic architecture. A packaged, versioned sub-agent that publishes a data contract to a separate optimization engine. Determinism is enforced as a test, the module boundary is checked at process level, and continuous integration gates its own published result counts so they cannot drift. On top of it I am architecting an agentic LLM framework with a knowledge ontology that makes cost-aware experimental decisions, choosing adaptively between a cheap observation and an expensive full characterisation inside a multi-fidelity loop.

Neural-guided search. Monte Carlo Tree Search and bidirectional A* over AND/OR graphs for synthesis planning, with a learned cost-to-go heuristic that cuts exploration at unchanged per-step compute. First-author conference abstract. See the projects →

Where it comes from

My PhD at Kyoto University was on making second-order information affordable at scale, and I pushed every method through to an application rather than stopping at the theorem. A regularized L-BFGS method in Computational Optimization and Applications. A Barzilai-Borwein variant of SVRG in JIMO. The regularized Nyström curvature line, which reduces per-iteration cost from \(O(nd^2 + d^3)\) to \(O(ndm + dm^2)\) with a sketch of roughly 5 to 20 percent of the feature dimension, and keeps the step a descent direction even when the Hessian is indefinite.

Those methods went into brain-MRI tumour classification with transfer learning, where data is scarce and getting more out of each gradient actually matters; into federated learning on edge devices, where the Hessian for a ResNet-50 does not fit on the client and a sketch takes the per-round cost from hours to minutes; and onto high-dimensional classification benchmarks and the CUTEst collection. More on the research →

How I work

Experiments designed before they are run. Paired statistical tests rather than single-run deltas. A stated detection floor beside every comparison. Negative results reported with their mechanism rather than dropped. Most of the levers you try do not work, and knowing which ones did, and why, is the job.

Day to day: Python, PyTorch, BoTorch and GPyTorch, NumPy and SciPy, pytest and CI, Git, MATLAB, RDKit. Things I have shipped: packaged libraries with versioned contracts, benchmark harnesses with thousands of tests, reproducibility checks that fail the build when a figure stops following from its data.

Currently interested in research scientist and research engineer roles where optimization and decision-making are the substance of the work: AI for science, autonomous experimentation, industrial R&D, efficient training, and scientific computing.