Hi Nathaniel,
I know we are right at the finish line for GSoC 2026 submissions! I’ve been working intensely on a highly practical, memory-efficient architecture for the AI-Assisted Log Diagnosis & Root-Cause Detection project, and I wanted to get your eyes on it before the final deadline on Tuesday.
Instead of a heavy, memory-intensive approach, my proposal focuses on a lightweight CLI pipeline:
-
Chunked Data Extraction: Using
pymavlinkto read.binfiles in 10-second rolling windows so the tool doesn’t cause OOM errors on massive 100MB+ logs. -
Lightweight Inference: Training an XGBoost or 1D-CNN model on common failure modes (Vibration, GPS Glitch, etc.) and exporting it to ONNX for fast, dependency-free execution.
-
Fix Retriever: A module that maps the model’s predicted anomaly directly to ArduPilot Wiki documentation for actionable user feedback.
My background is in building end-to-end ML architectures. I’ve recently deployed an Agentic AI system (Dark Pattern Detective) on Hugging Face, and built an OpenCV system translating human eye-blinks to Morse code—which taught me how to extract accurate patterns from noisy, time-series sensor data.
Here is the link to my complete proposal and timeline: GSoC_2026_ArduPilot_Proposal - Google Docs
Since time is short, could you take a quick 2-minute glance to see if there are any major architectural red flags or ArduPilot-specific constraints I missed? If the pseudo-code and timeline look solid to you, I will lock it in for final submission.
Thank you for your time and for mentoring this amazing project!
Best regards,
Mohid
GitHub: B-Mohid (B_Mohid) · GitHub