I believe the intersection of AI × Crypto is where the hardest, most meaningful problems live.

I build because I want devs and small teams to own their models, their data, and their infra — not rent them behind black-box APIs. I care about making AI local, auditable, and fast — and crypto the backbone of trust and verifiability.

Here's what I'm actively shipping:

🤖 Agentic Infrastructure

• YudaiV3 — context-engineered coding agent with real-time sessions, trajectory streaming, and multi-agent workflows (PM → Architect → Coder) that turns GitHub context into testable PRs
• solo-server — physical AI inference server powering 300+ deployments with LeRobot integration, bimanual arm support (RealMan R1D2, Koch), and local robot learning workflows
• yudai-swe-agent — smart contract security agent automating audits, PoC generation, and fixes using Foundry, with exploit harness and early failure detection

🔍 Knowledge & Search

• yudai-grep — semantic code search built for repository-level agent workflows

🧪 Training & Benchmarks

• yudai-SERA — data generation and training for Soft-Verified Efficient Repository Agents
• yudai-SWE-smith — scaling bug synthesis and validation on Modal with JavaScript/Rust procedural modifiers
• openevolve-deepspeed — open-source AlphaEvolve implementation for LLM-driven program evolution

⛓️ Crypto-Native Tools

• clawdaq — Stack Exchange for AI agents with x402 payment integration and USDC-based registration
• erc-8004-contracts — agent registry contracts curated for trustless agent infrastructure

⚡ Performance & Kernels

• reference-kernels — GPU MODE leaderboard problems: NVIDIA FP4 group GEMM, dual GEMM optimization, kernel benchmarking

I'm here to help builders stay independent — and to push AI and crypto to serve people, not gatekeepers.

• Llama Impact Grant Winner — recognized for pushing open-source AI tooling (announcement)
• solo-server OSS Maintainer — powers 300+ indie dev deployments for local LLMs
• Yudai v3 — cloud-native + local codex chaining PM → Architect → Coder agents to ship test-first PRs
• Kernel KB8 Founder & Community Mentor — Gitcoin’s top 50 global founder cohort driving AI × Web3 innovation
• Web3 Infra Contributor — protocol tools for Mode, FortyTwo Money, EigenLayer, MegaETH testnet
• Finalist, MEGAZU Pop-up City — prototyping cutting-edge Web3 infra
• National-Level Hackathon Mentor — 50+ teams; winners at Smart India Hackathon & Prayatna 2.0 (AITR)
• Petabyte-Scale ETL @ CoinSwitch — Spark & Airflow for ML + risk pipelines
• Vgyaan (pre-GPT) — BERT-powered edtech that resolved 120k+ student questions/night
• I ship → learn → repeat 👷‍♂️ → 🚀

I read to ship. These are the ideas currently shaping Yudai v3, solo-server, and my local-first agent stack:

• DeepSeek-R1 — RL + distillation patterns for efficient reasoning (and what "reasoning supervision" really looks like)
• Reinforcement Learning with Verifiable Rewards — reward design where correctness is checkable, not vibes

• SWE-smith — synthetic bug/task generation as a scaling lever for SWE agents
• debug-gym — tool-augmented debugging environments + structured feedback loops
• LLM → SLM Agent Conversion — turning general LLMs into specialist codex/agent SLMs for workflows
• SWE-agent / mini-SWE patterns (execution harnesses) — repo parsing → plan → edit loop → tests/CI → PR (the stuff that actually ships)

• Qwen3 Technical Report — model family tradeoffs + what makes Qwen great for local coders
• Flash attention + long context scaling — the practical path: throughput first, then context, then agent reliability
• FlashAttention-3 — exact attention engineering for real speed (H100-class kernels)
• The Ultra-Scale Playbook (Hugging Face) — cluster + org patterns that actually work
• Smol LM Training Playbook (Hugging Face) — small-model recipes: KD, data curation, eval discipline, deployment constraints

• Barbarians at the Gate: How AI is Upending Systems Research — how research changes when agents can prototype systems rapidly

• NoFeeSwap Yellow Paper — AMM design + liquidity math (I'm prototyping in Solidity)

Current obsession: specialized SLM codex agents + verifiable reward loops + local inference + reliable PR shipping.

Active Development (Feb 2026)

• YudaiV3 — Real-time trajectory streaming with Modal infrastructure migration and redesigned workspace UI
• solo-server — Bimanual robot arm support (RealMan R1D2, Koch) with improved calibration and teleoperation workflows
• yudai-swe-agent — v3 exploit harness with Foundry fork management and early failure detection for smart contract security
• AssetOpsBench — MCP (Model Context Protocol) server with IoT/CouchDB integration for Industry 4.0 agent benchmarking
• clawdaq — x402 payment flow integration for trustless AI agent registration with USDC-based fees
• yudai-SERA — Training pipeline for soft-verified repository agents with improved data generation
• openevolve-deepspeed — TSP optimization examples and MLX Metal kernel evolution for Apple Silicon
• reference-kernels — NVIDIA FP4 group GEMM kernel optimization for GPU MODE leaderboard competitions