alejandrozarzuelo.com

Alejandro Zarzuelo Urdiales

AI+math/physics researcher building open-problem maps, auditable technical systems, and original mathematical research.

I was the lead scientist at Archivara, where my work helped create the technical story behind its YC S26 path and reported $25M valuation. Across AI mathematics and physics I have produced 32 papers so far, including the first semi-original Erdos Problem 897 result, the first and so far only Epoch Math problem solved with AI, a 16 x 16 matrix multiplication result, and clique-cluster graph theory. My current work includes the 8,785-record Ulam Open Problem Difficulty Atlas, CostFloor, and the MeloText-ID symbolic-music benchmark. I am broadly multimodal and adaptable across STEM fields, using AI wherever there is real value to create.

Early research milestones

Turning frontier-model capability into work mathematics could inspect.

I entered AI-assisted open-problem research before it had a settled workflow. My contribution was not simply prompting a model: it was deciding which questions were tractable and consequential, directing the search, checking the mathematics, establishing what was genuinely new, and communicating the result under expert scrutiny.

First semi-original Erdos Problem 897 result with AI

I led the research process behind an AI-produced counterexample to Erdos Problem 897, from selecting the question and running the search to verifying the argument and explaining it to the mathematical community. The result mattered because it moved beyond closed-form exercises: the system produced a complete argument on a live research question, and the work then had to withstand public scrutiny about correctness, provenance, and originality.

Read the Erdos 897 paper

First and so far only Epoch Math problem solved with AI

As lead scientist at Archivara, I was the researcher behind the first and so far only AI solution to an Epoch Math open problem. The significance is not the label alone: it showed that careful problem selection, model direction, and mathematical review could convert frontier-model capability into a result outside the distribution of standard benchmark questions.

Epoch Math post

First bountied problem solved with AI

Resolved Clark Kimberling's Special Numbers problem 23 through a result on C-finite subsequences of Beatty sequences. Because the question carried a public bounty and a concrete mathematical criterion, it offered an unusually clear test of whether the research workflow could deliver an externally legible result.

Bountied problem post

Defended results with leading mathematicians

I have explained and defended results in public discussions with Terence Tao and other mathematicians, including work on the Optimization Constants list. This is part of the job that model output cannot replace: stating the claim precisely, locating its limits, responding to objections, and revising the presentation when scrutiny reveals ambiguity.

Optimization Constants post

A 32-paper research record across mathematics and physics

The breadth of the portfolio is evidence of a repeatable research process rather than one fortunate result. Across optimization, graph theory, combinatorial geometry, planetary science, symbolic physics, and model evaluation, I have repeatedly learned a new technical setting, identified a tractable contribution, and carried it through to a public artifact.

Browse the research corpus

Lead scientist at Archivara

Lead scientist for the AI-for-science research program whose technical work helped support Archivara's YC S26 trajectory and reported $25M valuation.

Multimodal STEM adaptability

I adapt quickly across mathematics, physics, spatial reasoning, optimization, scientific workflows, and any STEM field where AI can create value.

San Francisco technical ecosystem

Active participant in San Francisco tech events such as Frontier Research Club, and happy to join recommended events where I can meet people in the industry or present my work.