Systems that define our craft
We focus on projects at the intersection of AI, finance, and infrastructure — systems where engineering depth creates real competitive advantage.
LexRAG — Legal Contract & Compliance Intelligence
Enterprise document intelligence platform deployed across DIFC and ADGM for law and accounting firms. Multi-format ingestion (PDF, DOCX, scanned), local hybrid RAG for clause extraction, regulatory cross-referencing, and structured risk reports — under 3 minutes per document at 94% accuracy.
pico-type & NanoForecast: Edge AI Foundation Models
We designed, trained, and open-sourced pico-type (a 1.5M param byte-level text classifier running under 6ms on CPU) and NanoForecast (a zero-shot time-series forecasting model from 200K params compatible with Raspberry Pi). Deployed inside air-gapped systems for absolute data privacy.
TinyDoc-VLM: CPU-runnable Document Vision Model
A 256M-parameter vision-language model optimized for processing high-volume documents privately on CPU. Deployed inside air-gapped corporate setups to perform OCR, layout parsing, and visual Q&A without cloud dependencies or GPU costs.
fugusashi: Federated LLM Routing Engine
An intelligent, privacy-preserving LLM routing middleware. Uses federated learning and human-interpretable routing tables to direct LLM requests across model tiers, saving 40–70% in API costs while keeping sensitive queries on-premises.
Multi-Asset Quantitative Trading Platform
Production-grade quantitative trading system with multi-timeframe analysis, Smart Money Concepts strategy engine, live exchange execution across BTC, ETH, SOL and XRP. Dynamic position sizing, trailing stop-loss, and real-time PnL reconciliation with zero manual intervention.
AI-Powered Market Intelligence Terminal
Intelligence terminal with contrarian edge detection engine comparing model consensus against prediction market prices. Real-time feeds, sector heatmaps, and multi-model ensemble analysis for actionable trading opportunities across financial and event domains.
KARN — A Programming Language for AI Agents
We designed and built KARN from first principles — a token-minimal, platform-agnostic language that compiles to C, JavaScript, and WASM from a single source file. Native interop with pip, npm, cargo, and system libs. 4× denser than Python with error-as-value, async-by-default, and three execution modes.
Have a project in mind?
We take on select projects where engineering depth creates real impact.
Start a Conversation →