From smart contracts to small language models.

A 58M-parameter French reasoning LLM trained entirely from scratch — on a gaming laptop's RTX 4060. It works through problems in a <think> scratchpad before answering, and beats public French GPT-2s twice its size on language quality and arithmetic. Qwen-style architecture, Muon optimizer, custom digit-split BPE tokenizer, full pretrain → midtrain → SFT → GRPO/RLAIF pipeline.

A native desktop app to run LLMs entirely on your own machine — no cloud, no API keys, no data leaving the device. Inference through llama.cpp with CUDA, Vulkan and Metal backends, an agentic system with 30+ tools (files, shell, git, web search), MCP support, granular permissions and a bilingual FR/EN interface.

An agent harness with one door, and a receipt for everything that goes through it. Most harnesses ask the model not to leak your credentials — locked removes the ability. Every third-party call leaves through TAP, the sandbox has no network interface at all, and the compiler enforces that the tool executor can't even open a socket. Prompt injection can fully persuade the model — and still find no tool to comply with.

Built with the Etherwave Labs team for Funding the Commons' RealFi hackathon — and it won first place ($5,000) in the human.tech track. An event platform that kills the no-show problem: attendees lock a USDC deposit in a trustless escrow contract, check in by QR code on event day, and the contract automatically redistributes absent participants' deposits to those who showed up.