I build and experiment with AI systems that do iterative technical research: forming hypotheses, writing implementations, running experiments, analyzing failures, and optimizing against external evaluators.
My current project is Active Model, an experimental research system built around that loop.
Two independent #1 results on NVIDIA's official SOL-ExecBench leaderboards, evaluated on NVIDIA B200 hardware.
Active Model first optimized SOL-ExecBench kernel 049 and independently produced two candidate kernels whose measured performance would have placed them #2 on the public leaderboard at the time. I did not publish either result.
The remaining gap was difficult: the existing #1 was a strong solution that the initial campaign could not crack. I therefore added an external reasoning loop around Active Model.
I repeatedly pasted the current best Active Model solution and its measurements into ChatGPT, using Astra at maximum reasoning effort, and asked for new candidate architectures, variations, and optimization directions. I then passed those proposals back into Active Model, which converted them into actual kernel implementations, ran them on B200, measured them, analyzed the results, and continued the search.
The loop was repeated several times:
Active Model kernel → B200 measurements → ChatGPT/Astra research proposals → Active Model implementation and experimentation → new best kernel
That iterative model-to-model research loop eventually broke through the existing #1 and produced the kernel that now holds #1 on NVIDIA's official B200 leaderboard.
Leaderboard · Full campaign writeup
Autonomy boundary. Active Model independently reached approximately #2-level performance before external steering. Throughout the subsequent campaign, Active Model remained the implementation and experimental engine: it produced the actual kernels, executed the benchmarks, interpreted measurements, and incorporated successful ideas. ChatGPT/Astra was used between rounds as an external research adviser to propose new directions from the current solution and measured evidence. I orchestrated the exchange between the two systems.
The implementation is currently withheld while the benchmark remains competitive; the leaderboard result is NVIDIA's measurement.
Active Model holds #1 on NVIDIA SOL-ExecBench kernel 094 (time_decay_exponential_stabilization) with a 0.999092 SOL score on NVIDIA B200, ahead of submissions from SF Tensor, doubleAI and Databricks.
Leaderboard · Full campaign writeup
Autonomy boundary. Active Model produced every implementation, experiment, measurement and attribution in this campaign. A human research lead set strategy and, after the first measured round, named specific techniques to try. The model implemented and adjudicated 13 such proposals against its own measurements — adopting 3 and refuting 10 with identified mechanisms — and independently found the largest single discrepancy of the campaign: a reproducibility discrepancy inherited from prior work, traced to its underlying mechanism.
The first 4 autonomous hours reached a locally measured ≈#8; model-generated engineering under human research direction reached #1.
The first #1 (0.998647) was overtaken days later by another participant at 0.999005. A second campaign, started from the current kernel with the first campaign's work as read-only input, produced on its own a substantially different kernel architecture that overcame a limitation the first campaign had treated as fundamental and retook the lead on its first official measurement, and three rounds of written review widened the margin to +87 ppm over the participant who had overtaken it: 43 B200 rentals, 12.2 GPU-hours, $45, reproduced across four official evaluations. While the benchmark remains contested, the techniques are withheld; the result is NVIDIA's measurement.
-
llama.cpp — PR #27478 (open): "ggml : speed up batch-1 CPU decode, align large allocations". Up to +15.29% end-to-end token-generation throughput on Ryzen 7 9700X (attention change alone +10.67%) and +9.22% on Neoverse-N1, measured on Qwen3-30B-A3B Q4_K_M at 8192-token context.
-
zstd — three optimization PRs on encode/decode hot paths submitted upstream: #4729 · #4732 · #4733. Includes a decompression-hot-loop optimization removing a loop-carried memory dependency, improving decode throughput by +2.7–3.4% with GCC and +5.5–9.3% with Clang on Zen 5, with the same direction on Intel Raptor Lake.
-
dav1d — two AV1 decoder optimizations contributed upstream. !1967 refmvs: collapse runs of identical temporal MVs, merged September 23, 2026: +0.5–0.8% overall decode performance across x86-64 and AArch64, up to +2.96% on favorable content. !1968 mc: avg_direct for full-pel compound blocks, submitted upstream: +0.9–2.8% decode performance depending on architecture and existing SIMD coverage. Both preserve bit-identical decoded output.
- Solved the English perfect 10-square, a problem open for more than 120 years — repository · problem history.
- Improved published Google DeepMind AlphaEvolve results on packing problems using the published verifier.
- New best-known numerical results in circle packing and spherical codes (S⁵, N=86 and N=98, registered with spherical-codes.org).
- Improved the published primal for the 13,659-bus European Pegase AC optimal-power-flow benchmark; accepted into MINLPLib.
- WA — phenotype-driven gene prioritization: a Bayesian diagnostic model that ranks candidate disease genes from patient phenotypes alone, using the Human Phenotype Ontology. Preprint included in the repository.
How far can frontier models go when given a real objective, an evaluator, tools, compute, and enough freedom to conduct experiments?
Interests: autonomous research · AI for systems · model self-improvement · GPU/CPU optimization · LLM training and inference · algorithm discovery · verifier-guided search
I am also interested in applying the same verifier-guided search approach to rare-disease diagnostics and drug discovery — problem classes with concrete objectives, external validation, and search spaces far too large to explore by hand.
I am open to research-engineering and research roles at frontier-model labs where this kind of work is relevant.