Systems Architect · AI/IoT Engineer · Founder of Asirive Singapore · "Failing with Honour."

Cypher mascot, co-engineer, resident fox

About me I'm Haziq. I bridge high-level multimodal AI with low-level edge hardware. I build accessible, high-performance systems: assistive wearables, an in-house edge language model, a decentralized triage engine, on-device vision apps, and competition infrastructure. I document failures openly, including a "renovation graveyard" of post-mortems on designs that didn't work. Bugs are necessary renovations for character. I won the TKKYIA 2026 Merit Award with a team of three from Admiralty Secondary, where I led the systems side.

I founded Asirive to build ambient software and assistive hardware. Two products are the core of it, and the team page lists the people behind it.

Cortex is spoken navigation for blind and partially sighted cane users, co-designed with the Singapore Association for the Visually Handicapped (SAVH). It is built around one idea: the white cane already handles the ground, so Cortex stays quiet about that and only speaks up about what a cane can't catch.

Cortex V3: dual-camera glasses and the Raspberry Pi 5 pocket unit, joined by a single USB-C cable	Worn: a lighter, open-ear build
Field testing with an SAVH participant, white cane in hand

How it evolved, one round of user feedback at a time

Design decisions that matter

• Cane-aware safety: it only warns about overhead hazards and fast-moving objects, never ground clutter the cane already finds
• Reflexes stay local: hazard detection (YOLO on the Hailo accelerator) runs on-device in under 100 ms, with no network
• Two voices: conversation and critical safety warnings use different TTS voices, switchable within 100 ms, so you can tell them apart by ear
• The cloud only when it helps: Gemini Live for conversation, plus routing and bus arrival times, with memory synced in the background
• Tested with the people who use it: a year of SAVH field sessions (2025 to 2026)

Repo · Product page

Aether is an in-house edge LLM (REV 0.1, in development, formerly SNAP-C1), trained from scratch on a single consumer GPU.

• About 187M parameters with a custom architecture mixing PolyMamba2 (state-space) and CC-FSA (frequency attention), plus Mixture-of-Depths routing
• Trained and run on one AMD RX 7600 (8 GB) using ROCm/HIP on Windows 11
• About 26.30 BF16 TFLOPS sustained, roughly 3,900 to 4,200 training tokens per second
• 6.55B-token training target, 116 passing tests

Repo · Product page

Cyph, the Furry Engineer started as a terminal pet for a robot arm and grew into a small system: a personal co-engineer with a body. It lives on my desktop and my phone, plans my day with me, remembers me, talks, and moves a real 6-joint arm (VectorM, below).

_{The desktop app. Screens here are the real UI running on sample data, not my own calendar or chats.}

The phone app	The Arm page (mock camera and arm agent)
Widget on my real home screen	A reminder card, with his face as the avatar

flowchart LR
    You([You]) --> Desk["Desktop app"] & Phone["Phone app<br/>widgets, notifications"] & TG["Telegram"]
    Desk & Phone & TG --> Core["Cyph Core<br/>FastAPI, pinned TLS<br/>chat, plan, memory"]
    Core <--> Gem["Gemini"]
    Core --- Mem[("Memory + planner DB")]
    Agent["Arm agent (PC)"] -- "outbound websocket" --> Core
    Agent --> Arm{{"VectorM arm"}}

Privacy is a design rule: memory, transcripts and keys stay on my machines, Telegram is owner-only, and widgets never show chat text on the lock screen. Study mode, phone voice, QR pairing and push notifications are not built yet, and the repo says so plainly.

Cypher's mood reel

VectorM is the body: bare-metal firmware for a 6-DOF robot arm. I replaced the stock firmware on its STM32 controller so the PC can do the thinking (kinematics, planning, vision) while the chip delivers precise, glitch-free servo motion over a CRC-checked serial protocol. It works on its own too: anything that speaks the protocol can move the arm.

from arm_client import ArmClient

with ArmClient("COM15") as arm:
    arm.move_joints([135, 135, 135, 135, 135, 135], time_ms=1500)
    arm.stream_trajectory(waypoints, dt_ms=20)    # path playback
    arm.torque(False)                             # go limp to pose it by hand

VectorM repo · Cyph repo

AI / ML	MCP quantized edge models
Hardware	Elegoo Neptune 4 Pro parametric CAD logic analysis I2C / SPI / UART
Firmware and systems	device drivers
Software	WebRTC Web Audio
Tooling	Claude Code OpenCode and other autonomous terminal agents

• TKKYIA 2026 Merit Award (Tan Kah Kee Young Inventors' Award): the highest award in the Student Category that year, shared with a team of three; I led systems
• A year of SAVH field sessions (2025 to 2026) co-designing and user-testing Cortex with the people who use it
• Build with Gemini XPRIZE 2026: Asirive Copartner entered before the competition retired
• HackerOne researcher: irsplays

"I view bugs not as failures, but as necessary renovations for character."