David Balan — Systems & AI

About

I'm a 3rd-year Computer Science student at Queen's University specialising in AI — currently on exchange at the University of Manchester. I think in systems: how components compose, where performance is lost, and how software meets the physical world.

My work spans embedded firmware, computer vision pipelines, and full-stack applications — with a focus on correctness, efficiency, and real-world impact over theoretical elegance.

Dean's List 2024–25. Mayor's Award at KingHacks 2026. Two-time VEX Robotics World Championship competitor. Firmware shipped at Neuronicworks.

Currently

Semester abroad at University of Manchester · Jan – June 2026

English FLUENT

Romanian FLUENT

Russian LEARNING

EDU_01

Queen's University, Kingston B.Computing (Honours) — AI Specialisation · GPA 3.72

EDU_02

University of Manchester Semester Abroad · Jan 2026 – June 2026

STACK

Primary languages C++, Python, TypeScript — and the tools around them

AWARD

Recognition Dean's List · Mayor's Award, KingHacks 2026 · VEX Build Award, Worlds 2025

Projects

Selected work.

★ Featured Project

CleanShare

Desktop app that detects and blurs sensitive content — alcohol, licence plates — from images using a custom-trained YOLOv11 model and a hybrid C++/Python inference pipeline. Trained on ~1,000 annotated images; processes 1080p in ~5s on CPU-only hardware.

C++ Qt OpenCV YOLOv11 ONNX Runtime CMake

~80% detection accuracy · 1080p in ~5s on CPU · packaged for Windows

~1k

Training images

~80%

Detection accuracy

~5s

Per 1080p image · CPU only

01 — FEATURED

CleanShare

Desktop application that detects and blurs sensitive content (alcohol, licence plates) from images using YOLOv11 and ONNX Runtime. Trained on ~1,000 annotated images. Processes 1080p images in ~5 seconds on CPU-only hardware.

C++ Qt OpenCV YOLOv11 ONNX Runtime CMake Inno Setup

TECHNICAL CHALLENGE Discovered PyTorch → ONNX export inconsistencies mid-project. Implemented a hybrid C++/Python bridge: images pipe from C++ to a Python inference script, bounding boxes return via a temp JSON file — restoring reliable detections without rewriting the Qt pipeline.

~80% detection accuracy · 1080p in ~5s on CPU · packaged for Windows deployment

What I'd do differently

I'd investigate the PyTorch → ONNX pipeline earlier — the hybrid C++/Python bridge worked, but a cleaner solution would be to run inference fully in Python as a sidecar process from the start, with a defined IPC contract.

← move away to close

$ ./cleanshare --model yolov11n.onnx

Initialising ONNX runtime...

✓ Model loaded · trained on 1,000 imgs

C++ → Python bridge active...

✓ Hybrid pipeline ready

img_001.jpg | 4.8s | detections: 2

img_002.jpg | 4.9s | alcohol: 1, plate: 1

img_003.jpg | 5.1s | detections: 0

CoffeeBot Firmware

Firmware for an ESP32 microcontroller monitoring coffee machine usage via an amp sensor. Samples every second, transmits aggregated state over Wi-Fi to a WLAN server every ~5 seconds — networked displays render real-time usage.

C++ ESP32 I²C Wi-Fi State Machine

TECHNICAL CHALLENGE Tight memory constraints from the display footprint on the ESP32 — optimised buffer usage and eliminated dynamic allocations to achieve a stable system with near-zero incorrect state displays.

Deployed at Neuronicworks · stable 24/7 operation

What I'd do differently

I'd add a lightweight watchdog timer from day one and instrument the state machine with telemetry — debugging edge cases on a deployed ESP32 without visibility is painful.

← move away to close

Kingston CareMap

Low-data mobile app for Kingston helping users find essential services and food resources under real constraints — limited connectivity, low battery, time pressure. Dual navigation: interactive map + list fallback with need-based filtering.

React FastAPI Python MySQL Docker

TECHNICAL CHALLENGE Designing a friction-reduction flow that minimises dead ends from stale data — prioritised verified listings, displayed last-updated timestamps, and flagged uncertain details with explicit warnings.

Mayor's Award, KingHacks 2026 — top civic-impact prize

What I'd do differently

I'd invest more time in caching the service data locally so the app is fully functional offline — the dual-navigation fallback helped, but true offline-first would have made it far more resilient in the field.

← move away to close

FaceCrypt

Locally-running biometric encryption tool that uses face detection as a gate to protect sensitive files. Simple Tkinter GUI manages encrypted files and the unlock flow — from idea to demo in 36 hours.

Python OpenCV face_recognition Tkinter AES

TECHNICAL CHALLENGE Deriving a stable, deterministic encryption key from a high-dimensional face embedding that remains consistent under varying lighting and pose conditions.

Built in 36h · presented to judges at hackathon

What I'd do differently

The key derivation from the face embedding was brittle — I'd use a proper fuzzy commitment scheme or a dedicated biometric key derivation library rather than hand-rolling the embedding-to-key mapping.

← move away to close

Experience

Where I've worked.

May – July 2025

Firmware Developer Intern

Neuronicworks Inc. · Toronto, ON

Implemented firmware features enabling reliable comms between displays, sensors, and embedded controllers
Conducted QA testing across projects, documenting integration issues and collaborating to resolve them per SOPs
Created detailed test cases and validation reports to reduce debugging turnaround time

Sept 2025 – Present

Teaching Assistant — CISC 204

Queen's University · Logic for Computer Science

Led office hours clarifying course concepts and providing targeted assignment feedback
Supported 50+ students bi-weekly, improving assignment scores by ~10% on average
Collaborated with instructors to standardise grading and ensure accessible learning support

Sept 2023 – Present

Strategy Lead & CV/AI Integration

Queen's VEX Robotics Team · Dallas, TX — World Championship

Trained and integrated YOLOv11 object detection models for real-time game-state perception
Led Strategy team — analysed competitors and match data to develop match-winning approaches
Competed at 2024 & 2025 VEX Robotics World Championship; won Build Award, Worlds 2025
Acted as liaison between the robotics team and Queen's School of Computing faculty

David
Balan.

Selected work.

CleanShare

CoffeeBot Firmware

Kingston CareMap

FaceCrypt

Where I've worked.

Technical stack.

What I can talk to.

Let's talk.

DavidBalan.

Selected work.

CleanShare

CoffeeBot Firmware

Kingston CareMap

FaceCrypt

Where I've worked.

Technical stack.

What I can talk to.

Let's talk.

David
Balan.