Lisa Welsch

Native code intelligence as a single binary. Deep code understanding locally via CLI and MCP — semantic search, call graphs, impact analysis, architecture mapping. Zero LLM costs, zero latency.

Structural intelligence engine for codebases. Maps dependency graphs, detects architectural bottlenecks via Bridgeness Centrality, enforces layer boundaries, scores codebase health (0–100), and simulates the ripple effects of changes before you make them. Rust core with MCP server and BM25 search.

Language-agnostic open protocol for streaming, incremental code intelligence. Bridges LSP's runtime-query model with SCIP's static-snapshot model — adding a lazy query graph, blast-radius indexing, and a content-addressed dependency registry. Rust reference impl with Dart bindings.

Automated architecture audits in the CI/CD pipeline. Detects pattern violations, circular dependencies, and structural drift before they become technical debt.

Interactive onboarding system that radically shortens time-to-productivity for new developers by teaching project-specific architecture and conventions.

High-efficiency RAG pipeline. Transforms 25+ formats (PDF, audio, video) into token-optimized knowledge stores. Massively reduces context load while maximizing retrieval precision.

Institutional memory for AI-assisted teams. Preserves tacit project knowledge and human insights across context resets and personnel changes. CLI + MCP server.

High-efficiency context compression for LLMs. Reduces token usage and inference cost while increasing relevance and semantic fidelity.

Zero-dependency TypeScript library for splitting structured formats (JSON, XML, YAML, Markdown) into parts that must survive verbatim and parts that can be compressed. The seam-splitting logic extracted from ContextCompressionEngine as a reusable package. 87 tests. Works in Node, Deno, Bun, and edge runtimes.

Orchestration framework for complex multi-agent systems and swarm intelligence workflows within knowledge processing pipelines.

Adaptive serving layer that detects hardware capabilities (CPU, RAM, GPU) in real-time, enabling servers to deliver tailored content (lite vs. full) per device.

Capability-centric, local-first LLM routing layer. One line of config between your app and your models — routes each prompt to the cheapest capable option, including on-device models. Tracks cost, caches patterns, learns over time. 445 tests.

LLM-native pixel art toolchain. Parse, validate, render, and generate tilesets from .pax files — a TOML-based format designed for how LLMs actually think. CLI, MCP server, HTTP API, and Flutter desktop studio.

Open bioinformatics platform for cancer research. Autonomous AI agents producing cited gene dossiers and variant interpretation.

Genome analysis, LLM-powered evidence synthesis, and whole-cell simulation with CellForge.

Reference implementation of Morphogenic Intelligence — neural networks that grow, differentiate, and restructure at runtime without backpropagation. Solves CartPole-v1, reaches 87.7% stateless MNIST accuracy, self-heals after 30% neuron damage. Python and WASM bindings.

Pixel-art puzzle game for iOS. Swipe controls, progressive difficulty, original soundtrack.

Free, open-source retro pixel-art city building simulation. Custom isometric rendering engine, mod-focused architecture, cross-platform.

Recipe management app for iOS and Android. Offline-first architecture, Firebase sync, web scraping imports, collaborative sharing.