8
4
New Zealand
1 year of experience
I specialize in the intersection of civil infrastructure and artificial intelligence. I build and deploy production-ready AI tools that solve complex engineering and operational challenges. My recent deployments include computer vision pipelines for AMDS-compliant road and powerline assessments, as well as enterprise RAG systems and LLM-powered business intelligence dashboards. Core Competencies: • Domain: Civil & Transportation Engineering, Infrastructure Asset Management. • AI & Development: Computer Vision (ONNX, GPT-4o), Agentic AI (LangGraph, OpenClaw), Local LLMs (Ollama), Hybrid RAG, Python, Streamlit, Azure OpenAI. • Spatial Data: Creator of LidarLens (3D LiDAR point cloud processing package). • Technical Communication: Translator of O'Reilly technical books on 3D Data Science and AI Agents. I am passionate about turning complex infrastructure data into streamlined, automated digital solutions.

In high-precision electronics and micro-assembly manufacturing, surface-mount technology (SMT) defects—such as solder bridges, skewed QFP chips, and soldermask micro-fractures—cause catastrophic field failures if undetected. Traditional automated optical inspection (AOI) machines are brittle, rigid, and disconnected from robotic manipulation. **AMAI** closes the perception-action-supervision loop: 1. **Simulated Physical AI Workcell:** A calibrated PyBullet digital twin with an overhead RGB-D camera and a 6-DoF SO-ARM100 robotic manipulator with parallel jaw gripper. 2. **Procedural Defect Synthesizer:** Real-time generation of realistic PCB substrates, IC chip packages, copper traces, and injected micro-defects (solder bridges, skew tolerances, fractures). 3. **Intel OpenVINO Edge Vision:** Deep neural defect detection running at **0.24 ms** per board with 2D-to-3D projection into world coordinates. 4. **VLA Motion Planning:** Collision-free grasp affordance reasoning avoiding keep-out zones and generating smooth multi-waypoint trajectories to sort units into Pass vs. Scrap/Rework bins. 5. **Speechmatics Voice Supervision:** Real-time speech streaming allowing human operators on the factory floor to issue verbal overrides, freeze lines during safety events, or flag recurring defects hands-free. 6. **Immutable QA Audit Trail:** Persistent SQLite and JSON audit logging capturing every unit's telemetry, visual defect classification, confidence, 3D coordinates, and voice transcripts.
16 Sep 2026

1. Mission & Real-World Impact Construction is one of the most hazardous industries globally, accounting for over 1 in 5 workplace fatalities annually. The OSHA Focus Four hazards (Falls, Struck-By, Caught-In/Between, and Electrical) represent over 60% of all construction fatalities. 2. The Solution: SafeSite-AI deploys a hybrid, two-stage Perception + Multimodal Reasoning Agent directly onto on-site edge hardware (such as laptops equipped with AMD Ryzen™ AI or AMD Radeon™ GPUs): Stage 1 (Edge Perception): Local YOLOv11n detects personnel, heavy equipment, and scaffolds in milliseconds while computing spatial proximity vectors and danger zone geometries. Stage 2 (Local Multimodal Reasoning): Conditioned spatial prompts are dispatched to AMD Lemonade SDK running locally at http://localhost:13305/v1 for zero-latency hazard reasoning. Autonomous Agent Loop: The agent invokes local deterministic tools to search OSHA standards, audit PPE checklists, calculate machinery swing clear zones, and compile formal inspection reports. 3. Creativity Offline Edge Autonomy: Transforms static vision models into an active, multi-turn AI Safety Officer capable of dispatching edge alerts and authoring formal audit reports with 0ms cloud dependency.
1 Dec 2026

In New Zealand, over 600,000 households (nearly 35% of the population) live in residential rental properties. When an emergency strikes—such as a burst hot water cylinder, a gas leak, or a broken heat pump in winter—delays in reaching property managers cause tens of thousands of dollars in water and structural damage, while exposing landlords to Tenancy Tribunal fines of up to $7,200 under the Residential Tenancies Act 1986 (RTA) and Healthy Homes Standards. KiaOra Dispatch is an enterprise-grade autonomous Voice AI Property Maintenance Coordinator designed specifically for New Zealand's property ecosystem. Built on the AssemblyAI Voice Agent API and powered by Universal-3 Pro, KiaOra Dispatch fields inbound tenant phone calls 24/7, conducts natural spoken conversations with full barge-in support, understands Kiwi terminology ('heat pump', 'toby valve', 'chattels'), and automatically classifies issues across statutory priority tiers (P1 Urgent Emergency <4h SLA, P2 Healthy Homes Breach <24-48h SLA, P3 Routine Maintenance). Mid-dialogue, the agent executes autonomous function calls (dispatch_work_order), assigns certified emergency trade contractors (such as Auckland Master Plumbers and Laser Electrical NZ), delivers immediate life-safety and water-isolation directives, and dispatches structured JSON work orders directly to nationwide agency webhooks including Barfoot & Thompson, Harcourts, Ray White, and Quinovic. Architected with Next.js 14 App Router, Web Audio API (24kHz 16-bit PCM downsampling), and server-side ephemeral token minting to protect permanent API keys, KiaOra Dispatch reduces emergency dispatch latency from hours down to seconds, dropping call handling costs from $35 to under $0.45 per call. By combining AssemblyAI voice intelligence with codified tenancy law, KiaOra Dispatch turns speech into instant, legally compliant property action.
30 Sep 2026

VibeHedge is an autonomous, cloud-ready AI trading and portfolio hedging agent designed to safeguard equity during severe market downturns. Built for the Alpaca AI Trading Agents Hackathon, the system monitors a $100,000 paper trading portfolio and enforces an uncompromising 2.5% drawdown risk gate ($97,500 capital floor). When equity breaches the threshold, VibeHedge executes an integrated three-tier defensive protocol: Risk Assessment: FinRL-X continuously monitors portfolio state, Sharpe ratios, and live drawdowns against the high-water baseline. Downtrend Forecasting: ForecastAgent ingests multi-asset hourly market data and 34 FinRL technical features to predict forward macro trajectory and evaluate downtrend probability using deep sequence models. Quantitative Options Structuring: Vibe-Trading's Options Lab computes analytical Greeks (Delta, Gamma, Vega, Theta) across live option chains, structuring an optimal protective put (Delta ~ -0.35) constrained to a strict 1.5% portfolio cost budget. Execution is handled autonomously via a containerized FastMCP server implementing Server-Sent Events (SSE) and HTTP transports on Google Cloud Run. The agent connects directly to Alpaca's Trading API, Market Data API, and Model Context Protocol tooling, enabling ful
4 Sep 2026