The State of Robotics in 2025: Why the Hype Isn't Lying (Yet)
A deep dive into the four critical bottlenecks slowing the robotics revolution and why general-purpose robots remain inevitable
Jonathan Muhire
I build intelligent systems that ship.
ML systems, data infrastructure, and applied AI.
Selected Work
Research and systems I've built.
IBD scRNA-seq Benchmarking
Donor-aware benchmarking of CFN, linear, and VAE model families on two independent IBD cohorts. Strict leakage-safe cross-validation. AUROC 0.978. Targeting bioRxiv.
Reproducible Biomedical Eval Framework
Protocol-locked comparison of CFN vs XGBoost/MLP/logistic on NHANES hypertension data. Leakage guardrails, fold validation, and automated expansion gates.
Agent Evaluation — Memory Ablations
ReAct agent with memory-mode ablations and needle-in-haystack benchmarks. Eval-driven development: write the eval before the agent can pass it.
Vesuvius Challenge — 3D Segmentation
3D volumetric deep learning for detecting ancient papyrus surfaces in CT scans of carbonized Herculaneum scrolls. Topology-aware loss, ensemble + TTA.
RenAIssance — GSoC 2025
AI pipeline for digitizing Renaissance-era manuscripts using LayoutLMv3 layout recognition and Tesseract OCR with post-correction.
Writing
Notes and breakdowns.
Making Sense of Multimodal Models with Partial Information Decomposition
A deep dive into how Partial Information Decomposition (PID) reveals how different modalities interact in AI systems, from redundancy to synergy