azri ehsan

01

Mīzān ميزان

Validates every verse in the Quran. Won't mislabel authentic scripture.

2026

An NLP validation engine that catches hallucinated Quran verses in LLM output, guarding against a failure mode where models generate text that looks Arabic but isn't scripture.

6-stage Arabic normalization pipeline + greedy binary-search fabrication detection.
Ships four ways: npm library, PWA web app, Chrome extension (MV3), and a FastAPI microservice on Railway.
Bundles an LLM accuracy benchmark testing GPT / Claude / Gemini on 50 Quran prompts.

TypeScriptPythonFastAPI DockerChrome Ext MV3PWA

azrirefik/quran-checker azrirefik.github.io/quran-checker

02

MachineIQ

AI-powered predictive maintenance platform, real-time telemetry & alerting for industrial AHUs

2026 · Gamuda AI Academy C6 · in progress

An AI-powered predictive maintenance platform built at Gamuda AI Academy C6, for facilities managing 10-100+ industrial AHUs. Ingests real-time telemetry from IoT gateways, calculates health scores, detects anomalies via a configurable rule engine, and generates AI-powered repair guides from uploaded SOPs and repair history.

Configurable rule engine: flow-graph anomaly detection (trigger, condition, action), time-based escalation chains for technician notifications, sandboxed javascript conditions via quickjs.
FastAPI + LangGraph backend: 14 REST endpoints, 5 relational tables, event-driven telemetry ingestion, offline detection, AI knowledge base for step-by-step resolution guides.
React 19 frontend with TypeScript, shadcn/ui, 21-day telemetry visualizations (airflow, fan speed, filter dp, temp).

FastAPIPostgreSQLReact TypeScriptTailwind CSSDocker

code: in progress_ more on github

03

Insect Classification

Full Research Study, CNN architecture comparison & interpretability

2025

A complete ML research study comparing three CNN architectures for 5-class insect classification. The interesting finding: a model with 4x more parameters performed worse. Transfer learning won with only 3k trainable params.

Trained 3 CNN architectures on CIFAR-100 insects. surprising finding: deepcnn (391k params) performed worse than simplecnn (94k). resnet-18 transfer won with 81% and only ~3k trainable parameters.
Full ablation study: augmentation was worth +13%, batchnorm +9%, dropout actually hurt by 4%. augmentation matters more than model size.
Grad-CAM revealed custom CNNs looked at background color instead of the insect. transfer learning fixed that. documented in an academic paper.

PyTorchtorchvisionscikit-learn matplotlibLaTeX

code

04

PainPoint

Lovable x AWS Hackathon 2026

2026 · Healthcare Track

A dialect-aware health companion for elderly patients who don't speak standard English, with tap-to-report body map, voice input in Hokkien & Cantonese, caregiver dashboard with emergency SOS routing.

Dialect-responsive AI: Hokkien & Cantonese-aware symptom reporting for elderly users.
Tap-to-report body map + caregiver dashboard with emergency SOS routing (Leaflet).

React 19TypeScriptTanStack Start Tailwind CSSLeaflet

code

azri ehsan.

Selected work

Mīzān ميزان

MachineIQ

Insect Classification

PainPoint

Stack & tooling