Most AI repositories trending today are one-off prompt wrappers: they look impressive in 10-second cherry-picked screen recordings, but collapse under production load, network latency, or unexpected real-world edge cases.

Over the past few months, I set out to build and open-source four production-grade engineering systems designed for real-world stress, offline environments, and mathematical determinism.

Here is a breakdown of what I built, how each system works under the hood, and how you can run them today.

1. 🎙️ HardCall — 52ms Voice Interruption & Stress Cockpit

Building conversational voice agents is easy until you hit the interruption latency bottleneck. Human speakers naturally interrupt in ~200ms; traditional cloud STT/TTS chains lag at 800ms–1500ms, creating frustrating cross-talk.

HardCall is an enterprise-grade voice cockpit designed for ruthless behavioral stress-testing:

  • 52ms Barge-In SLA: Combines SenseVoice small acoustic models with raw FreeSWITCH RTP streaming to detect vocal intent and cut off agent playback in under 60ms.
  • Bi-Directional Full-Duplex: Zero-latency interruptibility with WebSocket and WebRTC audio pipeline.
  • Runnable Live Studio: Includes a browser cockpit simulating hostile HR, tech interviewers, and sales calls with post-call radar autopsies.

🔗 GitHub: github.com/wangguode/hardcall

🌐 Live Demo: hardcall.153557421.workers.dev

2. 🧩 Zhishu Agent — Self-Healing Coding Agent Harness

LLMs cannot reliably write correct software on the first try. Production agentic coding requires a harness architecture that isolates tool execution, validates state, and iteratively repairs broken tests.

Zhishu Agent implements an industrial agent loop pattern:

  • Self-Healing Test Loop: Runs unit tests inside sandboxed subprocesses, parses stderr stack traces, and prompts the agent with minimal diff contexts until all tests pass.
  • Dynamic Context Compression: Avoids LLM context bloat by pruning historical tool calls while preserving active AST invariants.
  • MCP Protocol Native: First-class support for Model Context Protocol servers and dynamic tool routing.

🔗 GitHub: github.com/wangguode/zhishu-agent

3. 📊 StatementFlow AI — Deterministic Double-Entry Financial Parser

Pure LLM extraction of bank statements and invoices is notoriously prone to arithmetic hallucinations (transposing digits, missing line items).

StatementFlow AI solves this with a dual-engine hybrid pipeline:

  • Semantic Extraction: Structured LLM extraction with strict Zod/JSON Schema boundaries.
  • TypeScript Deterministic Audit: Math is never computed by the LLM. An on-device TypeScript engine mathematically enforces accounting equality: Opening Balance + Total Credits - Total Debits = Closing Balance.
  • Client-Side Privacy: Financial documents are processed client-side without permanent cloud persistence.
  • Cost: ~$0.01 per document vs traditional $0.50 enterprise OCR fees.

🔗 GitHub: github.com/wangguode/statementflow-ai

🌐 Live Demo: statementflow-ai.153557421.workers.dev

4. 🌍 TERRA — 60FPS 3D Gaussian Splatting World Model OS

Travelers landing in a foreign airport face a chaotic "golden 120 minutes": SIM cards fail, airport Wi-Fi requires SMS verification, and offline 2D maps fail inside multi-level terminals.

TERRA is a spatial survival operating system:

  • 60FPS 3D Gaussian Splatting (3DGS): Photorealistic spatial digital twins rendered directly in WebGL on client devices.
  • Offline-First PWA: Add to home screen and run full spatial guides with zero cellular connectivity.
  • Multi-Airport Corridors: Production support for Seoul Incheon (ICN), Beijing Daxing (PKX), Tokyo Haneda (HND), and Bangkok Suvarnabhumi (BKK).

🔗 GitHub: github.com/wangguode/terra-world-travel

🌐 Live Demo: terra-world-travel.153557421.workers.dev

Summary & Full Source Code

All 4 projects are 100% open-source under permissive licenses. Every repo contains zero-configuration live demos and complete architectural blueprints.

⭐ Explore the entire portfolio on GitHub:

👉 https://github.com/wangguode

I would love to hear your feedback, architectural critiques, and thoughts in the comments below!