Javier Francisco
Dibo Gómez

Software engineer & edge-AI researcher

Currently completing a double master's — Computer Engineering and Cybersecurity — at the Universidad de Jaén. I build systems end-to-end and prefer them simple and correct.


Master's thesis · featured

A drone you fly by talking to it

In progress

You watch the drone's live feed and just say what you mean — "follow the white car," "switch to the motorbike," "look at that bird" — and the system turns the phrase into a bounding box, holds the lock through motion and occlusion, and drives the drone and gimbal to keep the target framed. Under the hood: a fine-tuned Qwen2-VL-2B, quantized to Q8_0 (1.65 GB) so the whole loop fits a 15 W Jetson Orin Nano with no cloud. A two-tier follow loop re-anchors with the VLM roughly every two seconds and coasts on a lightweight tracker in between, re-grounding on loss. I build it in pre-registered stages — single-frame grounding, then persistent tracking, then closing the loop to flight in simulation — each measured on held-out data at IoU@0.25, with the failures kept in the record rather than swept under it.

62%
language grounding on-device — beats the full-precision GPU reference
85%
tracking lock held through motion and occlusion
15 W
whole pipeline, on-device, no cloud
Read the repository →

Selected work