SilentWitness

Inspiration

SilentWitness was inspired by a hard reality: collecting evidence is often more dangerous than the abuse itself. Cloud-based apps leave traces—uploads, logs, notifications—that can be monitored or weaponized. In many abuse situations, internet access is unreliable or controlled by the abuser. We wanted to explore a different question: What if intelligence lived entirely on the device, and the cloud simply didn’t exist? SilentWitness was born from the belief that privacy should be guaranteed by architecture, not by promises.

What it does

SilentWitness is a covert, offline mobile AI that discreetly documents abusive or threatening situations. Disguised as a normal calculator app Runs 100% on-device with no internet access Listens only when activated Detects threat escalation and emotional distress Generates timestamped, legally structured evidence summaries Stores everything in an encrypted local vault Includes panic-wipe and decoy modes for user safety All data stays on the phone. Always.

How we built it

SilentWitness is designed around a local-first AI pipeline using the RunAnywhere SDK.

Architecture flow:

Audio Input → Quantized Whisper (On-device STT) → RunAnywhere Core (Orchestration) → DeepSeek R1 (Threat & Context Reasoning) → Legal Summary Generator → Encrypted Local Evidence Vault

Key choices:

Whisper (quantized) for robust offline speech-to-text DeepSeek R1 (distilled, 4-bit) for strong reasoning within mobile constraints RunAnywhere SDK for hardware-aware scheduling, memory control, and strict no-network enforcement The UI is intentionally minimal and disguised to reduce risk. Challenges we ran into Running reasoning-heavy models on mobile hardware required careful quantization and context control Balancing detection accuracy with user safety—false positives could be harmful Designing for worst-case scenarios, including no internet, device seizure, and hostile monitoring UX under stress: the app had to remain invisible, fast, and reliable in high-risk moments These constraints ruled out cloud AI entirely. Accomplishments that we're proud of

Designed an AI system that is impossible to build safely on the cloud

Created a privacy-first architecture suitable for life-critical scenarios

Demonstrated that on-device SLMs can perform serious reasoning tasks Aligned ethics, feasibility, and innovation into a single coherent concept SilentWitness proves that edge AI can be both powerful and responsible.

What we learned

Privacy is an engineering decision, not a legal checkbox Small Language Models are capable of meaningful reasoning when used correctly Offline-first design unlocks entirely new categories of applications In sensitive domains, less connectivity can mean more safety

What's next for SilentWitness

Expand multilingual and dialect support Improve threat detection with on-device personalization Integrate offline legal knowledge packs for different regions Conduct ethical reviews with survivor advocacy groups Explore partnerships for safe, responsible real-world deployment

SilentWitness is a vision for a future where AI protects users without ever watching them.

Built With

• deepseek
• runanywhere
• whisper