My Mission - Engineering the Unpredictable

From Safety-Critical to Context-Critical

For over a decade, I engineered safety-critical software for avionics and medical devices. In those industries, "mostly working" is a failure state. Systems must be robust, deterministic, and fail-safe.

When AI-assisted coding arrived, I observed how vibe-coding something from scratch can make impressive prototypes. I also observed how the best AI agents work on legacy codebases - applying wrong assumptions, making catastrophic mistakes, not seeing the whole picture - leading to architectural drift, context pollution, and technical debt. They build fast indeed - huge amounts of liability, if left unchecked.

This realization shifted my focus. I am building at the Harness Layer: the deterministic orientation substrates and architectural guardrails required to make AI agents safe, reliable, and economically viable for enterprise execution.

The Mission: Three Core Capabilities

I am developing three distinct capabilities at The Foundry. These are not just projects. They are infrastructure for the next generation of software architecture.

1. Codebase Orientation & Governance (repo-graph)

The Problem: Agents fail in real-world, legacy codebases because they lack structural orientation. They substitute hallucinated syntax for actual architectural awareness, leading to compounding technical debt and massive API token burn.

The Solution: A deterministic, multi-language codebase indexing substrate. It replaces grep and other tools with an explicit, queryable Abstract Syntax Tree (AST) graph accessed via command line.

The Goal: To constrain agentic entropy. By feeding models explicit boundaries, caller/callee edges, and policy gates before they generate syntax, repo-graph helps AI agents understand codebases better and stay on track.

Current work: https://github.com/andreirx/repo-graph

2. Knowledge Representation & The Three-Layer Truth Model (FRAKTAG)

The Problem: Standard RAG flattens wisdom into isolated text chunks, destroying structural hierarchy, ignoring uncertainty, and causing context pollution in expert-level reasoning.

The Solution: A fractal knowledge graph engine utilizing a biological-memory-inspired retrieval system and a strict three-layer truth model (extracted facts, unresolved observations, interpreted policy).

The Goal: To allow high-assurance enterprises to run localized, on-premise AI systems that act as grounded domain experts, completely isolated from data exfiltration risks and structurally prevented from generating "confident wrong" answers.

Current work: https://github.com/andreirx/FRAKTAG

3. Zero-Idle Serverless Infrastructure (AMODX & ZapEngine)

The Problem: The web infrastructure industry is trapped in monolithic technical debt, high idle costs, and vulnerable runtimes that stifle deployment velocity.

The Solution: Cloud-native, serverless operating systems (AWS Lambda/DynamoDB) and headless WebGPU/WASM data-oriented engines that physically bypass main-thread bottlenecks via zero-copy SharedArrayBuffer protocols.

The Goal: To provide deterministic, zero-idle-cost infrastructure that treats content as structured data, making it natively visible to the new generation of AI search engines and agentic workflows.

Current work: https://github.com/andreirx/zap-engine and https://github.com/andreirx/AMODX

The Philosophy

My work is driven by curiosity, solving difficult problems, and the idea of connecting high-fidelity minds with high-fidelity tools. The first generation of AI coding tools optimized for generation. The next generation must optimize for control.