Working systems with measured results: a planning agent for farm management that runs in the browser, a knee-MRI model for clinical imaging taken apart and reproduced, and a research plan for 3-D anatomical scene models.
A MuZero macrocontroller that plans each farm day over ten strategic plans, with a hemisphere crew layer and a Hungarian-assignment microcontroller below it. You can replay its games and re-run its tree search in the browser. T...
RSNA's 2026 knee-MRI challenge scores twelve findings by mean AUC. I took the 0.956/0.941 public tri-model blend apart down to tensor level and reproduced it exactly. I then tested three changes whose weights were fixed before ...
A research plan for a fine-tuned LLM that reads, repairs and writes OpenUSD (.usda) scenes. It is anchored on my Archetypal Knee USD model and scored with a layered metric, usdfid. The page covers the datasources and their lice...
All notebooks and datasets on Kaggle → · Toolkit: cheat sheet, Evidently template, Llama agents →
My notes on three talks from the graduate seminar.
One embedding can decide how much of a model, or which model, an input needs: MoE experts carved from a dense LLM, prompt-routed diffusion, cost-aware LLM routing and timestep-adaptive U-Nets.
How Realm, the runtime under Legion and cuPyNumeric, uses explicit events, a machine model and a DMA engine to overlap work, followed through a single linear layer.
Wristband data from older adults at home, contrastive multitask learning, and an impairment indicator that treats a classifier's mistakes as a cognitive signal.
Eight recent papers from AAAI, ILRT, IML and IEEE on LLM planning, formal verification and agent trace analysis, with notes on how each relates to verifiable-agent research.
A Trace-Language Framework for Agent Verification
Why textbook RSA preserves multiplicative structure and enables existential forgery, and how hash-then-sign and padding schemes like RSA-PSS close the gap.
How Lagrange and Cauchy shape real-world DH and ECC groups, why nearly-prime cofactor curves trade purity for performance, and what Secure Scuttlebutt teaches about validation.
From Weierstrass curves and double-and-add to cofactor attacks and Dual_EC_DRBG—how subgroup geometry and implementation choices determine ECC trust boundaries.
A structured look at HNP-based key recovery, BIGNUM representation leaks, lazy resizing CVEs, and protocol-level subgroup failures in modern signature stacks.
Bridging prime-order proofs with cofactor reality: small-subgroup attacks, BIGNUM leaks, twist security, and a practical mitigation checklist for protocol engineers.
Why finite-field DLP hits a scalability wall, how ECC and Pollard rho change the cost curve, and a worked finite-field point arithmetic example on F23.
A robust mathematical architecture unifying spatial statistics, duration-explicit sequence modeling, and Bayesian non-parametrics for sub...
A comprehensive hardware-software co-design project that simulates vLLM microarchitectural bottlenecks (KV cache, TLB) and applies those ...
A comprehensive reinforcement learning project implementing Q-Learning algorithm to teach an agent to play Pac-Man, demonstrating fundame...
Questions about cryptography, formal verification, or just want to connect? I'd love to hear from you.
A working notebook — half draft, half archive. The site indexes coursework, project write-ups, and longer pieces on cryptography and verification. Source on GitHub.