Systems & AI Engineer architecting autonomous agent runtimes, deterministic DAG workflow engines, and low-latency sandboxing infrastructure.
Website & Essays · RSS Feed · X (@abhiudayg) · LinkedIn · Dev.to
| System | Version | Role | Architecture & Stack | Repository |
|---|---|---|---|---|
agent-brain |
v1.0.0-rc3 |
Cognitive Runtime & Token Gate | Local-first cognitive governor, pre-tool route_gate hook, ~500-token knapsack packer, and agent-brain verify execution ledger (Rust, SQLite + HNSW, FastEmbed ONNX, stdio MCP) |
autonomic-ai-dev/agent-brain |
agent-spine |
v1.0.0-rc3 |
Workflow State Machine | Deterministic workflow DAG engine & cognitive state machine executing workflows/*.yaml with Tarjan SCC acyclicity checks and approval gates (Rust, Tokio, SQLite) |
autonomic-ai-dev/agent-spine |
dustmap |
v1.0.1 |
Native Systems Utility | High-performance macOS disk space analyzer and conservative cleaner with squarified treemaps, sector donuts, and parallel Rayon walk (Rust, Tauri 2, 100% offline) | abhiudayg/dustmap |
agentrun |
Pre-alpha |
Job & Container Runtime | Runs coding-agent tasks as isolated, recoverable jobs on your machine in dedicated containers and git worktrees with shared queues and audit trails (Go, Containers, Git Worktrees, macOS) | autonomic-ai-dev/agentrun |
harbor |
Pre-alpha |
Agent-Native Browser | Open-source agent-native macOS browser built on WebKit with explicit permission engines, human approval gates, SHA-256 hash-chained audit logs, and local MCP server (Swift, WebKit, macOS) | autonomic-ai-dev/harbor |
- agent-brain: the architectural pivot — from passive retrieval to deterministic cognitive runtime (Part 3)
Why passive context retrieval broke down in real-world autonomous coding, how we pivoted into a hard-gated execution runtime with token budgets, verify invariants, and upstream MCP routing. - agent-brain: memory scopes, temporal pruning, and cross-session routing (Part 2)
Multi-session context persistence, temporal fact decay, and circuit-broken upstream tools. - agent-brain: local MCP routing under a token budget (Part 1)
Answering the question: what should the model see right now under a strict 700-token ceiling? - agentrun: isolated coding agent runtimes with container workspaces
Designing hermetic, reproducible execution environments with copy-on-write workspaces. - dustmap: sub-12ms disk space visualizer in Rust and Tauri
Low-latency parallel directory traversal with Rayon and native macOS disk reporting.
- Verification as an Invariant: Every test, build, and claim requires cryptographic execution evidence (
agent-brain verify). Zero hallucinated success. - Token Budget Discipline: High-signal, byte-capped context envelopes over noisy prompt dumps.
- Zero Comment Slop: Self-documenting code with clear boundary discipline over LLM-generated narrative noise.
- Open Standards & Web Fidelity: Pure web standards, zero tracking pixels, zero analytics scripts, and 100% full-text RSS syndication.
- Languages & Runtimes: Rust, Go, Swift, TypeScript, Node.js, Bun, POSIX Shell
- Distributed & Systems: Tokio, SQLite + HNSW, FastEmbed ONNX, macOS WebKit, Container Sandboxing, Git Worktrees, Unix Domain Sockets
- AI & Autonomous Agents: Model Context Protocol (MCP), Agent-Spine Workflow DAGs, Cognitive Governors
- Web Platform: Astro, Tailwind CSS v4, Web Standards, KaTeX, Mermaid.js
Repository Source & Platform Build Details
This repository (abhiudayg/abhiudayg) also houses the source code for the personal brand platform and engineering blog (www.abhiudayg.dev):
# Clone repository
git clone https://github.com/abhiudayg/abhiudayg.git
cd abhiudayg
# Install dependencies
bun install
# Start local development server
bun dev# Run verified unit test suite
agent-brain verify -- bun test
# Run verified production build
agent-brain verify -- bun run buildLicensed under MIT. Written content published under CC BY 4.0.