"Failing with Honour."
I am Haziq (IRSPlays), a Systems Architect, AI/IoT Engineer, and Founder of Asirive based in Singapore. I specialize in bridging high-level multimodal AI reasoning with low-level edge hardware and ambient OS-level execution.
My engineering philosophy centers on building accessible, high-performance systems — from open-source assistive wearables tested with the Singapore Association for the Visually Handicapped (SAVH), to ambient OS-level AI harnesses for XPRIZE, to decentralized triage engines, computer vision iOS apps, and mission-critical competition infrastructure.
- Systems Architect & Founder @ Asirive - Building Asirive Cortex to tear down legacy hardware limits in assistive tech and ambient OS execution.
- Engineering Focus - Edge Vision AI, Custom Neural Architectures, Ambient OS Interaction, Real-Time WebRTC Systems, and Full-Stack Infrastructure.
- Vulnerability Research - Active security researcher on HackerOne (irsplays).
| Domain | Technologies & Tools |
|---|---|
| AI & Edge Computing |       |
| Hardware & Systems |      |
| Full-Stack & Web Systems |       |
| Cloud & Real-Time Data |     MCP Protocol WebRTC |
| Interactive UI & Audio |   Web-Audio Synthesizer Spatial Audio Cartesia TTS Supertonic |
| Project | Description & Core Highlights | Stack |
|---|---|---|
| Asirive Cortex | Hands-free AI wearable for navigation independence. ~90% cheaper ($156 BOM) than premium AI glasses. Tested with SAVH Singapore. Runs a 5-layer hybrid brain on RPi 5 (L0 Guardian offline safety <100ms, L1 Learner, L2 Thinker Gemini 3.1 Flash, L3 Guide NavEngine, L4 Memory). | Gemini 3.1 YOLO11n-NCNN Hailo-8L RPi 5 Python |
| Asirive Copartner | Ambient OS-level AI harness & Universal Semantic Bus. Built for the Build with Gemini XPRIZE. Eliminates mechanical execution friction via a 3-layer architecture (MCP, API Connectors, Computer Use) and a SNAP-C1 System 2 thought controller with ChromaDB memory. | Gemini 2.5 Python Rust / Tauri ChromaDB MCP |
| Competition-Management-Platform | Mission-critical robotics competition system for NRPC 2026. Complete platform handling 500+ teams with strict score calculation, poster submissions, and real-time live leaderboards. | React TypeScript Node.js SQLite Express |
| Project Echo | Decentralized AI-driven operations triage system. Replaces legacy school feedback pipelines with zero identity tracking and automated root-cause analysis. | Next.js TypeScript Python Tailwind CSS |
| SG60 Heritage Lens | On-device heritage recognition iOS application. SwiftUI app using a custom Core ML model to identify Singapore landmarks, food, transport, and cultural icons with interactive bite-sized historical context. | SwiftUI Core ML iOS Vision |
| SWAG Day FCS | Full custom stage AV & broadcast presentation system. Full-stack system featuring real-time controller/stage sync, procedural Web-Audio synthesis (zero audio files), and WebRTC audience phone camera broadcasting. | Next.js TypeScript Web-Audio API WebRTC Supabase |
| SNAP-C1 | From-scratch experimental neural architecture exploring self-improving code agents. V5 "Resonance" implements binary embeddings, spectral mixing, dual-speed learning, and an 8K elastic context window on consumer GPUs. | PyTorch Python AMD DirectML Custom Kernels |
| Portfolio OS (HaziqOS) | Interactive web operating system built as a developer portfolio. Features window management (drag, resize, focus, Z-index), 3D environment scene, Spotify clone with audio visualizer, and interactive CLI terminal. Live Demo → | React 18 Vite Three.js / R3F Framer Motion |
- Core UX Insight (Cane-Aware Safety): The white cane detects ground hazards. Cortex ignores ground clutter indoors and focuses exclusively on cane-invisible overhead hazards, fast-approaching objects, and visual info.
- 5-Layer Architecture:
- Layer 0 (Guardian): Safety-critical YOLO11n-NCNN + Hailo-8L depth (<100ms offline on RPi 5) with GPIO haptic feedback.
- Layer 1 (Learner): Adaptive open-vocabulary detection via YOLOE + MobileCLIP (~200ms).
- Layer 2 (Thinker): Scene reasoning and 14+ tool calling via Gemini 3.1 Flash Live (~500ms).
- Layer 3 (Guide): Intent routing, GPS turn-by-turn navigation, and LTA DataMall bus arrivals (<50ms).
- Layer 4 (Memory): Hot SQLite cache with Supabase 60s cold-storage sync (~1ms).
- Dual-Voice Priority TTS: Katie (female) for conversational chat, Troy (male) for critical safety warnings (<100ms distinguishing window).
- 3-Layer Interaction Engine:
- Layer 1 (MCP Protocol): Structured local filesystem, GitHub, and web actions.
- Layer 2 (Direct APIs): Connectors for Stripe invoicing, Vercel deployments, and Cloud Run hosting.
- Layer 3 (Computer Use): Screenshot capture + Gemini Vision + PyAutoGUI mouse/keyboard control for apps without APIs.
- Cognitive Hippocampus: 4-tier vector memory in ChromaDB (Semantic facts, Episodic task logs, Skill workflows, Preference profiles).
- Real-Time Stage Sync:
/controllerdrives/stagevia BroadcastChannel locally or Supabase Realtime cross-device. - Procedural Web-Audio Synthesizer: Zero static MP3 files — builds all audio beds, SFX pads, ducking, and crossfading programmatically in browser memory.
- WebRTC Live Stage Cam: Allows audience phones (
/audience) to stream live camera feeds directly onto the stage screen.
"I view bugs not as failures, but as necessary renovations for character."