Where AI Research Meets Bare Metal

Uday Rohith Reddy Yeruva trained in business analytics and now writes GPU kernels. His career has been built at the seam between how large models reason and how the machines underneath them actually execute — a seam that has become one of the field’s most contested pieces of ground.

There is a small, unglamorous operation buried inside modern AI systems called a fork. When a language model reasons through a problem by exploring several possible continuations at once — the technique that underpins tree search and much of what the industry now calls test-time compute — the system must duplicate the model’s working memory for each new branch. Do this often enough, at enough depth, and the copying itself becomes the bottleneck. The machine spends its time moving data rather than thinking.

Uday Rohith Reddy Yeruva spent a portion of the last two years attacking exactly that problem. His answer, a cache subsystem he calls FractalCache, replaced wholesale duplication with reference-counted sharing: branches point at the same underlying memory blocks until one of them actually diverges. On his benchmarks, using an eight-billion-parameter Llama model with a branching factor of eight, the fork cost fell from 36.7 milliseconds to 0.31 milliseconds. Peak memory dropped from 12.4 gigabytes to 2.6.

Those are the kinds of numbers that do not make headlines but quietly determine what is economically feasible to build.

An unusual route in

Yeruva’s path into frontier AI systems research is not the conventional one. He holds a bachelor’s degree in business analytics from Loyola Academy in Hyderabad and a master’s in business analytics from University College Dublin, completed in April 2026. Neither is a computer science doctorate, and neither is the credential typically held by people who write custom CUDA and Triton kernels for a living.

The pivot appears to have begun early. In 2021, while an undergraduate, he was selected through a competitive national research program for a data science research engineering internship at IBM in Hyderabad, where he worked on statistical modelling, high-performance optimisation and distributed machine learning pipelines. It was a first exposure to the infrastructure layer of applied machine learning — the plumbing rather than the model.

From August 2023 he worked as an independent AI systems researcher, based in Dublin and working remotely with collaborators in Mountain View, California. This is the period that defines his technical profile. He built post-training execution loops for PPO and GRPO — the reinforcement learning algorithms used to shape model behaviour after pre-training — handling custom loss formulation, gradient synchronisation and numerical stability across multi-node accelerator clusters. He designed asynchronous orchestration layers that decoupled prompt batching, cache allocation and reward verification so that expensive hardware sat idle less often. He reports cutting rollout iteration latency by roughly a quarter and halving experiment turnaround times.

That last figure is worth pausing on, because it explains the shape of everything that followed. Research velocity in modern AI is largely a function of how fast an experiment can be run and evaluated. Halving the cycle time doubles the number of ideas that can be tested.

Adversaries in the training loop

Two research efforts define his current work, and both concern what happens when the assumptions underlying distributed training stop holding.

The first, a preprint on Byzantine-robust decentralised GRPO, addresses a scenario that is increasingly plausible as training becomes distributed across parties who do not fully trust one another. If model post-training runs across independent nodes, what stops a malicious participant from poisoning the process? Yeruva identifies two attack surfaces: conventional domain poisoning, and what he terms reasoning-inflation denial of service — the observation that a bad actor can submit deliberately verbose completions whose inference cost inflates quadratically, degrading the system without producing obviously wrong outputs.

His proposed defence is a four-layer stack combining logit filtering, exponential-moving-average reputation scoring, Multi-Krum Byzantine fault-tolerant aggregation, and an adaptive penalty on completion length. He reports perfect reward recovery under a malicious fraction of 30 per cent, validated across seven cluster configurations with Welch t-tests and Cohen’s d effect sizes. It is, he claims, the first formalised framework for Byzantine robustness specifically in GRPO-based language model post-training.

The second, FlashReason, is under review at AAAI 2027. It combines a routing scheme for visual tokens with a modified policy optimisation method designed to separate perceptual errors from cognitive and symbolic ones — a diagnostic ambition as much as a performance one. The systems component is characteristic: a custom Triton kernel that dynamically routes decaying visual tokens out of the deeper transformer layers, reported to deliver a 1.88-times wall-clock speedup at a sequence length of 512.

A third project, Speculative Feature Steering, points in a different direction entirely — toward safety. It uses sparse autoencoders, a tool from mechanistic interpretability research, to intercept and steer a model’s internal activations in real time. The engineering trick is to fuse the entire encode-threshold-project pipeline into a single Triton kernel so the intervention is cheap enough to run continuously. Tested against 500 adversarial prompts on H100 GPUs, he reports suppressing targeted unsafe features by up to 88 per cent at a latency cost below 0.74 milliseconds per forward pass, with 97.1 per cent of general capability retained.

The common thread across all three is not the subject matter. It is the method: take an algorithmic idea that is expensive to run, then descend to the level of the kernel and make it cheap.

From clusters to control loops

Since 2026, Yeruva has held two founding roles simultaneously.

At New Atlantic in San Francisco, as founding lead systems and post-training researcher, he architects what he describes as a paper-to-environment pipeline: converting published scientific research in structural biology, chemistry and immunology into standardised reinforcement learning benchmarks for frontier AI laboratories. The work involves building hermetic containerised sandboxes around scientific tooling such as OpenMM and PyRosetta, so that training rollouts are reproducible and models cannot game the reward signal rather than solve the problem.

In April 2026 he founded Arthimic, Inc., splitting his time between Dublin and San Francisco, as its chief AI systems architect. The company’s focus marks a notable shift in his work. Where his earlier research optimised throughput across GPU clusters — how much a system can think — Arthimic concerns determinism: a native C++ control runtime holding a strict one-kilohertz execution contract for physical AI and robotics models. Every millisecond, without exception.

That constraint changes the engineering entirely. He has built cache-line-isolated, lock-free ring buffers to pass 100-byte frames between Python inference processes and the real-time control loop; quadratic-program barrier filters enforcing joint velocity limits on six-to-eight degree-of-freedom systems at microsecond-scale solver latencies, with fail-safe hold fallbacks; and copy-on-write background buffers so that model weights can be updated without ever blocking the control thread.

What the trajectory suggests

Read as a whole, Yeruva’s career describes a steady migration downward through the stack. He began with statistical modelling, moved to distributed training infrastructure, then to the kernels and memory systems beneath it, and now to hard real-time control where a missed deadline is not a slowdown but a failure.

It is also a career built substantially outside institutional research. Much of the work between 2023 and 2026 was done independently, published as preprints and open repositories rather than through a university lab. The field has, for now, remained porous enough to reward that.

Whether the two ventures he co-founded in 2026 endure is an open question. The problems they address — trustworthy decentralised training, and AI systems that must act in the physical world within fixed time budgets — are unlikely to be solved soon.

F A C T – C H E C K   N O T E S

The following claims derive solely from the CV supplied and would require independent verification before publication. Web searches conducted for this piece returned no publicly available corroboration for the items marked below; absence of search results is not evidence that a claim is false, but each should be confirmed directly.

Employers and roles

–     Arthimic, Inc. — no public record found (company registry listing, website, press coverage, or funding record). Confirm incorporation status in Ireland and/or the United States, founding date (stated as April 2026), and his title.

–     New Atlantic — no public record found. Confirm the organisation’s existence, his role and start date (stated only as “2026 – Present”).

–     Client attribution. The CV states this work produces RL benchmarks “for frontier AI labs for Anthropic.” The phrasing is ambiguous and names a third-party company. Any commercial relationship with Anthropic or other named laboratories must be confirmed with those parties directly. It is described in the article only as work for frontier AI laboratories, without naming clients.

–     IBM internship (Hyderabad, April 2021 – January 2022) — confirm with IBM. Also confirm the name and nature of the “competitive national research program” through which he says he was selected.

Publications

–     Flash Reason — stated as under review at AAAI 2027. Under-review status is not publicly verifiable, and papers under review are conventionally not reported as forthcoming. Request the submission ID or a copy of the manuscript.

–     Byzantine-Robust Decentralized GRPO (Dec-GRPO) — stated as a preprint on Research Square. No matching preprint located. Request a DOI or direct link. The claim that it is the “first formalized framework” of its kind should be treated as the author’s assertion; extensive prior literature exists on Byzantine-robust decentralized optimization more broadly.

Performance metrics

All quantitative figures — the 118× fork speedup and 4.8× memory reduction (Fractal Cache), the 1.88× speedup (Flash Reason), the reward margin of +0.052 and F₁ = 1.0 under 30 per cent adversarial corruption (Dec-GRPO), the 88 per cent feature suppression, 0.74 ms overhead and 97.1 per cent capability retention (SFS), and the 25 per cent latency reduction in his independent work — are self-reported and unaudited. None appear in a peer-reviewed venue. They are attributed to him in the text rather than stated as established fact. Ideally an independent reviewer would inspect the benchmark code at the cited repository, github.com/Uday Rohith-R/ml-systems-research, which was not publicly resolvable at the time of checking.

Education and dates

–     Confirm both degrees with Loyola Academy, Hyderabad, and University College Dublin.

–     Date inconsistency: the CV lists the bachelor’s degree as June 2021 – April 2024, but the IBM internship as April 2021 – January 2022, which begins before the degree. It also lists independent research work from August 2023, overlapping both the bachelor’s degree and the UCD master’s (September 2025 – April 2026).

–     Date inconsistency: two roles are listed as beginning in 2026 (New Atlantic, undated within the year; Arthimic, April 2026), overlapping the final months of the UCD master’s programme.

Media contact name: UDAY ROHITH REDDY YERUVA  

Organization: Arthimic, Inc. (Founder & Chief AI Systems Architect) / New Atlantic (Founding Lead Systems & Post-Training Researcher)  

Email: [email protected]

Website: github.com/Udayrohith-R

Phone: +353 892611905