I build the loops that make agents useful.
Agent orchestration, evaluation protocols, AEO, and production LLM features. Shipped in production and measured on public leaderboards. The open-source tools below are the proof.
Agent systems & orchestration
Multi-agent swarms over isolated git worktrees, MCP servers that expose the whole loop, role→model routing across providers with fail-closed adapters.
proof · #1 on ECDSA.fail. Bitte runtime in production, 2.85M+ agent messages
Evaluation & benchmark design
Anti-overfitting evaluation protocols, Pareto frontiers across competing objectives, keep/reject validity gating on every iteration.
proof · #1 on Optimization Arena's QEC decoder leaderboard
AEO / AI search
Answer Engine Optimization: AI-crawler robots policy, llms.txt and structured exports. I wrote the tooling that measures what ChatGPT, Claude, and Perplexity cite.
proof · creator of aeo.js + check.aeojs.org. 4,569 scans
Production LLM features
Streaming agent loops, OpenAPI→tool conversion, and on-chain agents that act, not just chat.
proof · ~20 on-chain primitives across NEAR / EVM / SUI / Cardano
Open Source
All AI tools →- Credential-theft detection
- Supply-chain analysis
The AI stack I actually use in production.
- Agent orchestration Race / consensus / pipeline swarms over isolated git worktrees: ralph-starter.
- MCP / tool use ralph-starter's MCP server exposes the whole loop to any MCP client.
- Evals & benchmarks Anti-overfitting protocols and Pareto frontiers. Autoresearcher's core loop.
- Multi-provider routing Role→model routing across 7+ LLM providers in the ECDSA.fail harness.
- Guardrails & spend gates Fail-closed adapters and budget gates on every autonomous run.
- Streaming UX Streaming agent loops in Bitte's production runtime.
- Structured output / function calling OpenAPI→tool conversion and function-calling manifests at Bitte.
- Memory Zep / GraphRAG-backed agent memory, productized at Mirofi.sh.
- RAG GraphRAG pipelines inside the multi-agent simulation engine behind Mirofi.sh.
- Vector search & embeddings The retrieval layer under the memory and RAG stacks above.
- Context engineering llms.txt + ai-index.json exports so LLMs read a site right. Shipped in aeo.js.
- Fine-tuning vs prompting Prompt + eval loops over fine-tunes. The bet behind both #1 finishes.
Ralph Starter
Specs drive code. AI handles the rest.
An open-source AI coding orchestration platform. Multi-agent swarm mode runs race, consensus, or pipeline strategies over isolated git worktrees; an MCP server exposes the whole loop to any MCP client; and a Figma→code visual validation pipeline closes the design loop. 100+ daily npm downloads.
- Multi-agent swarms: race · consensus · pipeline over isolated git worktrees
- MCP server: the full orchestration loop exposed to any MCP client
- Figma→code visual validation pipeline
- 5 spec sources: OpenSpec · Figma · GitHub · Linear · Notion
- Auto-runs tests, lint, build, captures errors shift-left
- 100+ daily npm downloads
Autoresearcher
General-purpose research loops.
A benchmark-driven autonomous research CLI. Multi-agent co-evolution: divergent agent populations explore in isolated git worktrees, champions merge back, and a Pareto frontier tracks the best candidates, with keep/reject validity gating on every iteration. Used internally at MultiVM Labs for post-quantum cryptography, smart wallet, and chain-level benchmarks, but intentionally general-purpose.
- Multi-agent co-evolution: divergent populations in git worktrees, champion merging
- Pareto frontier: best candidates tracked across competing objectives
- Keep/reject validity gating per benchmark metric
- Markdown final report + JSONL audit log per run
- Composable with any benchmark you can express as a shell command
AEO.js
Answer Engine Optimization for the modern web.
An open-source Answer Engine Optimization framework. Analyses your robots policy for AI crawlers and generates LLM-ready site exports (llms.txt, ai-index.json) so ChatGPT, Claude, Perplexity, and any LLM can discover and cite your site. Free, no signup, 100+ daily npm downloads.
- AI-crawler robots policy analysis: who can see what, and why
- LLM-ready exports: llms.txt · ai-index.json · per-page Markdown
- First-class plugins: Astro · Next.js · Vite · Nuxt · Angular · Webpack
- Standalone CLI mode
- Human/AI toggle widget, drop-in
- 100+ daily npm downloads
The agent fleet
The production side of this work: at Bitte I designed the agent pattern and runtime. An OpenAPI spec becomes a tool-calling agent with in-chat wallet signing. Then I shipped chain-specific agents on it that swap, stake, bet, and launch tokens for real.
I built the AEO layer. Now I run it as a service.
I'm the creator of aeo.js (100+ daily npm downloads) and check.aeojs.org (2,259 unique sites scanned). I know exactly what makes ChatGPT, Claude, and Perplexity cite a site, because I wrote the tooling that measures it.
AEO
A fixed-scope engagement: I audit your site against the answer engines, ship the structured-data / llms.txt / content-layer fixes, and hand you a before/after score report from check.aeojs.org.
Before / after report
Scored on the same rubric my scanner uses
You get the baseline scan, the shipped diff, and the re-scan. The improvement in your AI search score is measured, not asserted.
Fix your AI search score
Slots are limited. The work is hands-on, not delegated.
The ecosystem around the tools.
Templates, ideas trackers, curated awesome-lists, and shared cryptography tooling packages.
Hosted SaaS productized from an open-source multi-agent social-simulation engine (GraphRAG/Zep, OASIS). Auth, Supabase storage, Railway/Vercel deployment, web platform.
mirofi.shFive deployed chain agents, sole author: Solana assistant, Aerodrome (Base, ~25 tools), Morpho (Ethereum), meme.cooking (NEAR), Sui explorer.
github.com/bitteprotocolReady-to-use project templates consumed by ralph-starter.
github.com/multivmlabs/ralph-templatesPublic roadmap and changelog for the Ralph ecosystem.
github.com/multivmlabs/ralph-ideasCurated Ralph ecosystem links.
github.com/multivmlabs/awesome-ralphPost-quantum cryptography tooling for Rust and TypeScript: ML-KEM, ML-DSA, SLH-DSA, Falcon.
github.com/multivmlabs/post-quantum-packagesWant to integrate one of these into your stack?
I'm happy to chat. These are open source, but I'll often help teams adopt them.