AaronRockMenezes

I build grounded systems — ones that actually have to deal with the physical world or a real model’s internals, not just clean benchmarks.

The path looks scattered if you read it as bio→ML→robotics. Read it as predict → simulate → ground, and it’s the same line:

• Predict (BITS / APPCAIR / Deep Forest) — built models that had to be right about physical systems. Residue-level RNA-protein interaction prediction. Novel drug candidates and dielectric materials, validated by MD + DFT, not just held-out F1.
• Simulate (DeepChem, open source) — wrote the infrastructure that lets other people do that work. GPU neural-ODE solvers, segmentation pipelines for live microscopy, 7+ models shipped to 40k+ users. The work I’m proudest of from those years isn’t a paper, it’s that someone else’s drug-discovery experiment ran faster because of code I wrote.

• Ground (Brown / Serre Lab, now) — close the loop. Build embodied agents that have to act in 3D environments and reason about what other agents can see. Vision models still can’t reliably do line-of-sight reasoning — a faculty children develop by age 1–2. That gap is an AI safety problem (autonomous systems acting on a fake model of what others perceive), a mechanistic interpretability problem (if it’s not line-of-sight, what is the model learning?), and a computational neuroscience problem (VPT is a clean developmental milestone).

  My main project, VPTnav, is a synthetic benchmark + Isaac Lab pipeline for testing exactly this; alongside it I’m building NavJEPA, a lightweight action-conditioned predictor that runs mental simulation in latent space (not pixels — pixels are noise we don’t care about; what matters is where things are and what they occlude).

In parallel I tinker on the other side of the grounding problem — opening up what models have already learned. Most recent: mindweather, an SAE-feature steering setup on Gemma 3 (writeup). Same instinct as VPTnav, different direction: don’t trust the model’s outputs, look at its internals.

I’m currently looking for PhD positions and research opportunities — embodied AI, world models, mechanistic interpretability, vision. Drawn to groups that build the things they study. If that’s you, reach out.

Outside research: Valorant and Elden Ring, building physical things, good conversations about science or weird ideas.