DeepSeek vs. Llama

"Kill The Cloud" Challenge
Celebrating On-Device AI Innovation
🏆 Winner

OrbitMind

Team: ClauseZero
Platform: Unstop
Problem Statement

Autonomous navigation around non-cooperative Resident Space Objects (RSOs) is critical for future in-orbit servicing and space logistics missions, including refueling and active debris removal. These missions require a servicer spacecraft to operate at close proximity to targets that provide no cooperative aids, making the task inherently risky.

Solution

OrbitMind is an on-device autonomous navigation and docking system designed for safe interaction with non-cooperative RSOs. The system enables real-time pose estimation, navigation, and safety monitoring using only onboard sensors and computation with zero cloud reliance. Built with LLaMA for on-device intelligence, it provides lightweight, efficient reasoning optimized for spacecraft constraints with deterministic, fast inference critical for safety-critical control loops.

Team Members
Raghav Madan
Rishet Mehra
View Project →
🥈 Runner-Up

Beli

Team: Team Beli
Platform: Devpost
Problem Statement

In 2024, fraud shifted from hacking systems to hacking people. AI-generated voices and high-pressure tactics exploit human trust during phone calls that feel completely real. Cloud-based defenses introduce delays and privacy concerns that make real-time protection ineffective in moments of panic.

Solution

Beli (meaning "friend" in Punjabi) is an on-device firewall for human conversations. It runs silently during calls, analyzing speech locally to detect psychological manipulation patterns like false urgency, authority impersonation, and payment redirection requests. Using a dual-engine edge AI pipeline with on-device speech-to-text and reasoning models, Beli delivers discreet haptic vibration alerts without the caller knowing—all without internet, cloud, or data leaving the device.

Team Members
Aryan Saxena
View Project →
🥉 Second Runner-Up

SilentWitness

Team: SilentWitness
Platform: Devpost
Problem Statement

Collecting evidence in abuse situations is often more dangerous than the abuse itself. Cloud-based apps leave traces that can be monitored or weaponized by abusers. In many situations, internet access is unreliable or controlled, making cloud-dependent solutions inadequate for those who need protection most.

Solution

SilentWitness is a covert, offline mobile AI disguised as a calculator app that discreetly documents abusive or threatening situations. Running 100% on-device with no internet access, it detects threat escalation and emotional distress, generates timestamped legally-structured evidence summaries, and stores everything in an encrypted local vault. Built with the RunAnywhere SDK, quantized Whisper for speech-to-text, and DeepSeek R1 for threat reasoning, it includes panic-wipe and decoy modes for maximum user safety.

Team Members
Vijayalakshmi S
View Project →