Research

Papers

Long-form technical writing on the systems I build with. Each paper is also downloadable as a print-ready PDF (use your browser's Print → Save as PDF on the paper page).

Cryptography·14 min·2026-04

Post-Quantum Encryption for Production Systems

A Practical Playbook

NIST has finalized ML-KEM and ML-DSA. Industry still ships RSA-2048 in 2026. A production-tested migration playbook covering hybrid envelope encryption, signature replacement, and KMS plumbing — drawing on the @tigthor/kokocrypt deployment.

cryptographypost-quantumkyber
Read
AI Safety / Interpretability·15 min·2026-04

Inspecting the Loop

Mechanistic Interpretability for Tool-Using AI Agents

Mechanistic interpretability has spent three years on single-token forward passes. Production agents run loops with state spanning dozens of passes. Three primitives — trace factorization, decision-point attribution, counterfactual loop perturbation — for agent-loop interpretability.

ai-safetyinterpretabilityagents
Read
Cryptography / Blockchain·12 min·2026-04

Quantum Computing's Real Threat to Blockchain

A Migration Playbook

The 'quantum breaks Bitcoin' headline misses the actual threat. Hash functions are safe; ECDSA is exposed; address types matter enormously. A calibrated chain-by-chain analysis of the migration surface and a five-year timeline (2027-2032).

cryptographypost-quantumbitcoin
Read
Machine Learning·17 min·2026-03

Beyond Attention: The Future of Transformer Architectures

A 2026 Survey

Self-attention's O(n²) is not the wall — KV-cache memory is. A survey of the five families challenging the vanilla Transformer (linear attention, state-space models, hybrid stacks, MoE, Test-Time Training) and a prediction for the dominant 2027 architecture.

transformersmambastate-space
Read
AI Infrastructure·12 min·2026-03

Multi-Agent Coherence

Why Single-Player AI Tools Don't Scale to Teams

ChatGPT and Cursor are built for one human, one model. The next phase is multi-human, multi-agent coherence. Four primitives — shared context windows, witness-graph permissions, agent inheritance, CRDT state merge — for building team-AI products that don't fall apart at the seams.

ai-agentsmultiplayer-aib2b
Read
Distributed Systems·13 min·2026-03

BFT Consensus for Sub-Second Prediction Markets

Off-the-shelf BFT consensus (Tendermint, vanilla HotStuff) targets 1-12s blockchain finality. Prediction markets need 80-250ms. HotStuff-2 modifications: smaller validator sets, deterministic intra-block ordering, asymmetric trust, layered settlement.

distributed-systemsbftconsensus
Read
Game AI·13 min·2026-02

From Scripted to Sentient

A Framework for Persistent AI NPCs in Multiplayer Worlds

Generative AI in games is stuck in a 1.5 generation. A four-layer architecture (short-term memory + long-term semantic memory + social graph + goal stack) for NPCs that remember you across sessions and stay in character over hundreds of hours.

game-ainpcsllms
Read
Cryptography / Identity·11 min·2026-02

Zero-Knowledge Identity

On-Chain Credentials Without Exposure

The KYC-vs-privacy trade-off in DeFi is false. ZK proofs let users prove regulated predicates without revealing identity. What changed in 2024-2026 — PLONK-family provers, universal trusted setup, eIDAS 2.0 — that finally makes this deployable.

zk-proofsidentityplonk
Read
Blockchain / HFT·16 min·2026-01

Sub-100ms Execution: On-Chain HFT

Why HIP-4 ends the off-chain order-book era

Hyperliquid HIP-4 is the first architecture that resolves the AMM-vs-CEX dilemma: 80ms execution on a native CLOB, fully on-chain, with portfolio margin. The technical primitives that make it work and the implications for prediction markets, RWAs, and structured products.

blockchainhfthyperliquid
Read

9 papers · 123 minutes of reading · all original.

Open any paper, then ⌘ + P → Save as PDF for a print-ready version.