The Kairos Engine
A Cognitive Autopilot for BVLOS Mountain Drone Delivery
Built for Build With Gemma: Margadarshan, winning first place in the Gemma for Route Intelligence track. It gives a delivery drone something it has never had: judgment.
Read the full Kaggle writeupProblem
Today's delivery drones fly like trains on a track — when wind or battery fails mid-flight, the only fallback is Return-to-Launch. In the Himalayas that move is often fatal, because home may sit behind a 4,000 m ridge the drone cannot clear. The Kairos Engine is not a flight controller; it is a decision-making brain that sits on top of one.
Approach
Gemma 4 E2B runs the show — reading telemetry, calling the right tools (wind forecast, nearby safe landing spots, payload urgency), and looping sense → reason → act. An XGBoost model trained on 5,000 simulated Himalayan flights scores the crash odds for each move (fly home = 92%, divert = 5%) at 97.40% test accuracy. A weighted-Dijkstra pathfinder optimising energy, not distance, beats a naive shortest-path A* baseline.
Result
Mission Success Rate lifted across 20 random scenarios — each landing with over 15% battery to spare. The whole brain runs on the drone itself: Gemma shrunk to 4-bit (GGUF) on a llama.cpp runtime, even on a Raspberry Pi 5, so it keeps thinking when the ground link drops, re-planning in under 2 seconds. Every decision is logged, giving regulators (CAAN, FAA) the auditable trail BVLOS laws have been waiting for.
- Gemma 4 E2B agentic tool-calling
- Energy-aware weighted Dijkstra
- Runs on-drone (4-bit GGUF, Pi 5)
- Auditable decision logs for BVLOS
