Jony Jing

AI Engineer · Southern California

5 years shipping production AI systems — from RAG chatbots serving 3,000+ users to LLM evaluation pipelines to document intelligence platforms at 95%+ accuracy. Currently at Deloitte, building reliable GenAI for public-sector clients.

Emergence is not the enemy. Unauditable emergence is.

@jonyjing · LinkedIn · ClawHub

Current Focus

Engineering AI that can do more — expanding the inputs agents can perceive and the tools they can operate, while keeping everything auditable. Building skills and infrastructure in the OpenClaw ecosystem under a Scope-First framework: define the boundary, then let intelligence emerge freely within it.

Production Work

• RAG Policy Chatbot — Designed and deployed a RAG chatbot for a state government eligibility system. PoC → Pilot → Statewide (3,000+ caseworkers). Built the LLM evaluation pipeline (Ragas, 25 Golden Questions) now used as the client's standard for all model upgrades.
• LLM Code Generation Platform — GPT-4-powered natural language → Python/SQL for a federal financial department. Code preview, one-click execution, full audit trail for compliance.
• Document AI Accelerator — End-to-end OCR evaluation platform. Synthetic data pipeline (1,000+ docs, 3 languages, 81 handwriting fonts), 95%+ field-level accuracy.

ClawHub Skills

Published on clawhub.ai/JonathanJing:

• skill-trust-auditor — Security audit for ClawHub skills. Static analysis + LLM-as-judge with 4-layer prompt injection defense. Advisory-only: the LLM never affects the trust score.
• rag-eval — RAG evaluation toolkit built on the Ragas framework. Automated quality scoring for retrieval-augmented generation pipelines.
• openclaw-gateway-watchdog — Out-of-band gateway monitoring. Dual-layer: LaunchAgent + cron. Alerts even when OpenClaw is down. (GitHub)
• openclaw-dashboard — Analytics dashboard. Cache-aware cost estimation, mobile PWA, cookie auth. (GitHub)
• openclaw-tally — "Tokens tell you how much you paid. Tasks tell you what you got." Task-level efficiency analytics. (GitHub)
• glass2claw — Vision Router for smart glasses → Discord → Notion. Instruction-only design, 5 rounds of security hardening.

Church & Community — Grace-Irvine

Tech lead for Grace Irvine Church's digital infrastructure (volunteer):

• ProPresenter-Automation-Agent — AI agent that automates worship presentation workflows in ProPresenter
• Ministry-Scheduler — Automated scheduling for ministry teams
• Ministry-data-visualizer — Data visualization dashboard for church ministry metrics
• Ministry-Clean — Data cleaning pipeline for ministry records
• WorshipFlow — Worship service planning and automation
• AI-sermon-workflow — AI-assisted sermon preparation workflow
• catechism-cards — Interactive catechism study cards

Other Projects

• Asynchronous-FIFO — Async FIFO in Verilog ⭐38 — from my EE days, still getting stars
• Stock-Trading-Strategy — Quantitative trading strategy with backtesting ⭐7
• Web2Ebook — Scrape web content and convert to ebook format
• Hermes-handbag — Data analysis of Hermès handbag resale market ⭐1

Stack: Python · TypeScript · RAG · LangChain · Ragas · GCP (Vertex AI, Document AI) · AWS (SageMaker) · Docker · Spark/Databricks

Certs: AWS ML Specialty · Databricks GenAI Engineer · Microsoft Power BI