Alim Pradana

Operators point an Android phone at a shelf. A YOLO model counts the items in each area and pins their positions in 3D, so the next photo won't count them twice. Supervisors see the results on a web dashboard.

What it does

• Scan a shelf's QR code or barcode to set the location
• Real-time detection on the phone's GPU, about 35 ms per frame on a Pixel 6
• Operators can correct a count before they submit it
• Works offline and syncs when the phone reconnects
• Firestore rules separate operator and supervisor access
• CSV export and Google sign-in
• Unit and instrumented tests for the app

Built with

Kotlin, Jetpack Compose, ARCore, LiteRT (TensorFlow Lite), Room, Next.js, TypeScript, Firebase Auth, Cloud Firestore, Vercel

An internal REST API that counts bolts, nuts, gears, bearings, flanges, and long or short steel bars from a photo. I started it on YOLOv5, revised it twice, and later moved it to RF-DETR, a transformer-based detector that runs on ONNX Runtime. A Jetson Nano build with TensorRT FP16 ran at the edge from 2023 to 2025.

What it does

• Detection endpoints for a single image or a batch
• An endpoint that saves hand-corrected counts
• A training pipeline for custom datasets, with Roboflow for labeling
• Comet ML experiment tracking and hyperparameter tuning for each revision
• Stored results, so I can compare model versions over time

Built with

Python, FastAPI, Flask, YOLOv5, RF-DETR, ONNX Runtime, TensorRT, Roboflow, Comet ML, PostgreSQL, MySQL

A FastAPI service that runs on an edge device at the attendance kiosk. It registers workers' faces, matches each new photo with InsightFace, and checks the photo for spoofing. In a test on 5,432 past attendance photos it matched 99.2% of them to the right worker, averaging 879 ms per request with 30 requests arriving at once.

What it does

• Register several photos per worker and replace one at a time
• Match against all workers, or verify one worker ID
• Anti-spoofing with an ONNX model, with a heuristic fallback when no model is loaded
• An in-memory embedding index, so a match skips the full database scan
• API tokens with per-minute, per-hour and per-day limits
• A request queue that waits for free GPU memory, plus an admin endpoint that unloads the model
• Runs as a systemd service on Linux

Built with

Python, FastAPI, InsightFace, ONNX Runtime, CUDA, PostgreSQL, systemd

The system watches CCTV streams in real time and logs each time an employee uses a phone during work hours. It emails the alert list and exports reports as PDF. The company runs it in several branches.

What it does

• Start, stop and check a detection job for each camera and branch
• Violation logs grouped by location
• PDF report export
• Choose who receives alerts
• In-app and email notifications

Built with

Python, Object detection, RTSP, PDF reports, SMTP

Each location in the warehouse has a printed ArUco marker. The app stores item positions relative to the marker, so two different phones show an item in the same spot. I chose markers over pure detection here because this project needed precise locations.

What it does

• Scan a marker to identify the location
• See item positions in AR, anchored to the marker
• Admins manage locations and markers, users view and search items
• Audit log of data changes
• JWT sign-in with role-based access
• Database migrations with Alembic

Built with

Kotlin, Jetpack Compose, ARCore, CameraX, OpenCV ArUco, Retrofit, Room, FastAPI, PostgreSQL, SQLAlchemy, Alembic, JWT