Applied Mathematics & Computer Science · Universidad del Rosario · Bogotá, Colombia
Project guide · Updated CV: coming · Email
I build simulations and learning experiments, then test what their comparisons actually establish: strong baselines, held-out evaluation and explicit controls.
My research question is whether automated judges and monitors measure what they claim to measure. A pattern I keep running into: when a system is only partly observed, absolute levels are often not identified, while comparisons between systems can still be. Public work so far is on reinforcement learning, recurrent networks and human-coordination data. I am now applying the same tests to LLM judges and monitors (AI-safety audits, tutoring feedback) and to the METR time-horizon measurements. That work is in progress and not yet public.
Status: expected graduation 2028 (date to confirm). No publications yet. A short Motor-RNN note is being prepared for a preprint server; nothing has been submitted.
| Project | What I did |
|---|---|
| SCRES reinforcement learning · simulation |
Supervised project: supply-chain simulation and Gymnasium/PPO experiments judged against a same-contract static comparator. The learned policy was ahead in only 2 of 10 checkpoint means and 2 of 60 held-out streams; the earlier advantage claim was withdrawn. Start with the same-contract verdict. |
| Motor-RNN connectivity computational neuroscience |
Neuromatch team project, then my independent equal-plasticity control, which holds the number of trainable recurrent edges equal across densities. With all edges trainable, denser networks do better; with an equal trainable budget that advantage disappears (per-seed drop +0.231, 95% CI 0.193 to 0.267, positive in 16 of 16 seeds; it survives spectral-radius matching, random-subset budgets and N=400). One architecture, one task. Start with the robustness note (draft, not peer reviewed). |
| Spectral analysis of cognitive labor human coordination |
My reanalysis of Andrade-Lotero and Goldstone's human-search experiment (the data belong to the original authors), with a past-only temporal audit. Predictive advantage is not established. Start with the temporal audit. |
- SCRES: the PPO advantage over a strong same-contract static policy did not survive the comparison, so the earlier claim was withdrawn.
- Spectral: early-prediction AUCs of 0.804 and 0.860 used overlapping time windows and are not validated. No predictive or spectral advantage is established.
- Motor-RNN: with 8 seeds I reported that the density ordering reversed under an equal trainable budget. With 16 seeds that reversal is small and not robust (−0.041, 95% CI −0.083 to +0.011); what holds is that the advantage disappears. See the robustness note.
- HeliOS: a collaborative educational RISC-V kernel in C. I am a contributor, not the sole author. Architecture notes and QEMU smoke tests; smoke tests are not formal verification.
- ContratIA Abierta: open-procurement data pipelines, FastAPI services and traceable signals for human review, not automated allegations. Human validation is still pending.
- ChaosLab: double-pendulum dynamics with numerical checks and an interactive presentation. A physics course project, not an ML result.
- Merged pull request: ForagersEnv.step() simulation pipeline, in another group's repository (merged 2026-02-28).
- The Motor-RNN robustness note as a preprint.
- What measured anchors identify in METR's time horizons (preregistration drafted, not yet filed).
- Decision stability of LLM judges: how often a pass/fail or ranking decision changes when the judge changes, on AI-safety audits (Petri) and tutoring feedback.
- A report from the Apart Research collusion sprint (October 2026).
Described in the CV and not published: a retrieval-augmented support prototype with heuristic answer/escalation routing (synthetic data); numerical-analysis coursework; and a computational audit of a marmot social-network model, done with a collaborator whose name is withheld until they agree.
Read the question and the stated limitations, inspect a saved result and its method, then follow the documented test path. The existing evidence can be inspected without refitting any model.
Tools: Python · PyTorch · NumPy/SciPy · NetworkX · SimPy · Gymnasium · C · SQL · FastAPI · Git · pytest
AI use: I use AI coding assistants (Claude, Codex) for code and drafts. I check every reported number against saved outputs, and each repository states how it was built.