OvertakeGuard AI

What Inspired Us

In East Africa, two-lane undivided highways connect major economic hubs but claim thousands of lives each year in preventable head-on collisions. On corridors like the Addis Ababa–Jimma highway, slow-moving heavy freight trucks create massive blind zones. Passenger cars attempting to overtake are forced to pull blindly into the opposing lane—often colliding with high-speed oncoming vehicles like emergency ambulances and intercity buses.

This problem is deeply personal to our team: our lead engineer lost five family members on this very stretch of road during a graduation trip due to a blind-overtake collision behind a slow truck.

Cloud-based driver assistance cannot solve this reality. Rural highway corridors suffer from complete cellular network blackouts, and recurring API latency makes cloud calls deadly. OvertakeGuard AI was built to bring real-time, offline kinematic collision-prevention intelligence directly to low-cost, on-device computing hardware already accessible across Africa.

What We Built

OvertakeGuard AI is a zero-cloud, on-device vehicular safety advisor designed for resource-constrained edge hardware (8 GB RAM commodity laptops). Powered by a quantized 3-billion-parameter language model (Llama-3.2-3B-Instruct-Q4_K_M.gguf) running on pure CPU via llama.cpp, OvertakeGuard evaluates vehicle closing velocities, sightline occlusion dynamics, and provides split-second safety advisories in English and Amharic.

Key Capabilities:

• Zero-Cloud Offline Safety Engine: Evaluates kinematic overtaking safety margins with 0 internet connection.
• Bilingual African Road Guidance: Delivers immediate, high-priority maneuver instructions natively in Amharic (am) and English (en).
• Lean Hardware Footprint: Runs in ~3.4 GB RAM on standard CPU architectures without overheating or requiring expensive GPUs.

How We Built It

• Model Optimization: Quantized meta-llama/Llama-3.2-3B-Instruct to GGUF Q4_K_M, balancing complex trajectory reasoning with rapid token generation.
• Inference Runtime: Built on llama.cpp using OpenMP and AVX2 vector optimizations, with execution pinned to 4 physical cores to guarantee thermal stability below the 85°C challenge threshold.
• Domain Integration: Grounded in transportation safety engineering, time-to-collision ($TTC$) physics, and real-world African rural highway geometry.

Accomplishments We're Proud Of

• Validated complete offline functionality on Ubuntu 24.04 LTS on an Intel Core i5-8365U processor.
• Achieved 10.04 tokens/sec throughput ($S_{\text{perf}} = 66.93$) and a lean 3,438 MB RSS footprint ($S_{\text{eff}} = 50.88$).
• Preserved Ethiopic Unicode tokenization fidelity under 4-bit quantization for clear Amharic safety advisories.

What's Next

• Edge Camera Integration: Pairing the offline reasoning engine with lightweight on-device computer vision models to automatically calculate oncoming vehicle distances from an offset dashcam.
• Multi-Language Expansion: Adding localized audio alerts in Afaan Oromoo, Tigrinya, and Swahili.
• Fleet & Transport Deployment: Partnering with regional transport authorities to pilot offline edge safety units in public minibuses and commercial fleets.

Built With

• bash
• computer-vision
• edge-ai
• gguf
• llama.cpp
• llm
• python
• ubuntu