azri ehsan.

fullstack + ai · kl, malaysia

i build AI systems and ship them.

Selected work

01

MachineIQ

AI-powered maintenance intelligence platform for industrial AHUs, from early condition warning through guided resolution and maintenance learning

2026 · Gamuda AI Academy C6 · technical contributor + product owner / stakeholder-facing + QA contributor

Predictive maintenance intelligence for industrial AHUs. Early condition warnings, deterministic simulation evidence, and a RAG-driven resolution guide.

  • Quality-gated 7-day OLS regression with a supported filter signature: usable DP rising AND (usable CFM falling OR usable fan power rising); SAT can support evidence but cannot independently admit a warning.
  • Deterministic simulation proxy evidence: gradual degradation shows a 377h lead time, normal operation produces no false watch, sudden faults are rule-owned, and recovery clears the alert. This is simulation evidence, not field accuracy.
  • FastAPI / PostgreSQL / React stack with a RAG-based resolution-guide workflow over SOPs and repair history.
  • Technical contributor plus product owner / stakeholder-facing and QA contributor.
FastAPIPostgreSQLReact TypeScriptTailwind CSSDocker
code: in progress_ more on github
02

Mīzān ميزان

Quranic Verse Validation System

2026

Deterministic normalization validates every verse; LLMs are benchmark targets.

  • 6-stage Arabic normalization pipeline + greedy binary-search fabrication detection against 6,236 verified verses.
  • Ships four surfaces: npm library, PWA web app, Chrome extension (MV3), and a FastAPI microservice.
  • Bundles an LLM accuracy benchmark testing GPT / Claude / Gemini on Quran prompts.
TypeScriptPythonFastAPI DockerChrome Ext MV3PWA
azrirefik/mizan-checker azri.work/mizan-checker/
03

Insect Species Classification

Insect classification research study · CNN comparison & interpretability

2025

Three PyTorch approaches with custom CNNs and transfer learning.

  • Full ablation study: augmentation worth +13%, batch normalization +9%, dropout -4.4%.
  • Best test result 81.1%.
  • Grad-CAM revealed that custom CNNs relied on background color rather than insect features.
  • Documented in a 30-page academic paper.
PyTorchtorchvisionscikit-learn matplotlibLaTeX
code
04

PainPoint

48-hour hackathon · Lovable x AWS 2026

2026 · Healthcare Track

A dual-dashboard pain tracking app for elderly Malaysian patients and their caregivers.

  • Dual workflows: elderly patients tap a body map to report pain and log medications; caregivers track symptom history, book appointments, and route emergency SOS.
  • Voice input in five languages: English, Bahasa Malaysia, Mandarin, Cantonese, and Hokkien.
  • Stack: React 19, TypeScript, TanStack Start, Tailwind CSS, shadcn/ui, Leaflet.
React 19TypeScriptTanStack Start Tailwind CSSLeaflet
code

Stack & tooling

Languages
Python, JavaScript / TypeScript, SQL
AI / ML
PyTorch, NLP, computer vision, LangGraph, RAG, LLM APIs, Azure AI
Backend
FastAPI, PostgreSQL, SQLAlchemy, Docker, REST, JWT
Frontend
React, Tailwind CSS
Delivery / Tools
Git, Jira, pytest, Postman, GitHub Actions