Production systems, AI constraint engineering, and agent-operated infrastructure
Applied speech recognition research for Audio2SRT.app and production transcription systems
A product-facing ASR research track covering engine selection, word timestamps, speaker diarization, SRT quality, multilingual routing, and the gap between benchmark numbers and real production transcription output.
The Homoiconic Context: S-Expression Architecture Descriptors as Navigation Primitives for AI Coding Agents
Formal architecture descriptors (intent.lisp) as navigation substrate for AI coding agents. Controlled experiments show 33–44% step reduction (d=0.92) with architectural context, while S-expressions emerge as the only format without a fatal flaw across parsing, generation, and error resilience.
Neuro-Symbolic Code Generation via S-Expression Intermediate Representation and Deterministic Harness Engineering
A neuro-symbolic architecture that constrains LLMs to output S-expressions, validates against typed IR defined as Rust enum whitelists, and deterministically assembles guaranteed-compilable code. 75% Pass@1 compilation rate vs 62% for raw LLM generation.
An Operating System for AI Coding Agents: Process Management, Persistent Memory, and Vendor-Agnostic Orchestration
A Rust daemon (111K lines, 10 crates) that provides the missing OS layer for AI coding tools: slot-based process management, a growing knowledge base with hybrid search, semantic terminal parsing, and 71 MCP tools. Three independently publishable subsystems: Semantic Terminal, Decision Cascade, and Knowledge Flywheel.
Vendor-Agnostic AI Agent Observation Through PTY State Parsing
A multi-layer parsing pipeline that transforms raw PTY byte streams into finite state machines, enabling orchestration of any AI coding tool without custom APIs or vendor cooperation.
Graduated Autonomy: A Four-Tier Resolution Model for AI Coding Agents
A cascading decision architecture that routes AI agent questions through progressively expensive resolution tiers: KB lookup, cheap model consult, dedicated research session, and human escalation.
Persistent Cross-Session Memory for AI Coding Agents
A compounding knowledge system with hybrid search across typed entries, automatic experience extraction, retrospective learning loops, and budget-constrained context assembly.
Self-Evolving Microservice Cascade: A Four-Stroke Engine for Intent-Driven Autonomous Code Repair
A dual-loop automation engine: inner loop for deterministic cross-service repair, outer loop for repair-log distillation and template evolution.