Some people finish projects. I build things that don't need me to.

Started with one belief: a blind person shouldn't need a companion to move through the world. V1 ran on GPS/GSM and taught me web dev, Android, and backend from scratch — but it needed constant connectivity and had no intelligence. V3 rebuilt everything: Raspberry Pi 5, camera-based object detection, haptic feedback, and offline-first communication — no internet required, zero-downtime Kafka notifications on AWS.

V4 (in progress): turn-by-turn navigation, scene description, OCR, and fall detection.

Aerial object detection struggles with tiny objects, occlusion, and crowded scenes. We ensembled YOLOv8x and YOLOv8m using Weighted Boxes Fusion, tested on the VisDrone dataset — the combined model hit 0.45 mAP@0.5, beating either detector alone (0.431 / 0.400), balancing accuracy against real-time UAV constraints.

Backend engineer on a financial-crime compliance platform — multi-module Java/Spring Boot, Nuxt frontend, MSSQL at scale. Work spans authentication flows, load-testing infrastructure, and production debugging across a system built for regulated data.

Not trying to build the most powerful AI — trying to build one people talk to like a friend. A system with its own memory, that decides what to keep and what to let go. Currently simulating neurochemical responses and running a default-mode-network model that lets it wander, overthink, and follow its own line of thought.