Inspiration
Older adults are highly targeted by scams but often struggle with complex security apps. We wanted to build a "shield" (Kavach) right where they already communicate—iMessage and SMS—while keeping trusted family members instantly in the loop during high-risk situations.
What it does
Kavach is a community-powered AI scam-protection service.
Instant Checks: Forward a suspicious text or link to the agent for an instant verdict ("Looks safe," "Suspicious," or "Likely scam").
Family Alerts: High-risk threats trigger instant alerts to designated family members, requiring their acknowledgment.
Community Floor: Scams reported by 3+ users are globally flagged as a "Likely scam."
iOS App: A companion app offers a Call Companion, safety cards, and recovery checklists.
How we built it
We used SpacetimeDB (TypeScript) as the single source of truth for state and red-flag rules. The AI routing is handled by a Python/FastAPI service using the NeMo Agent Toolkit, NeMo Guardrails, and NVIDIA NIM (Nemotron 3.5 Lightning, Super VL, and Embed). A Node worker using the Photon Spectrum SDK manages all iMessage traffic. The client is a native Swift 6 / SwiftUI iOS app.To keep the experience natural, we optimized for low latency:$$T_{total} = t_{SMS} + t_{NIM} + t_{DB} \approx 5 \text{ to } 9 \text{ seconds}$$
Challenges we ran into
Preventing the AI from hallucinating "safe" verdicts was critical, requiring us to build strict, rule-based overrides for things like gift card requests. Orchestrating asynchronous SMS retry logic and family group threads alongside a strict database state was also highly complex.
Accomplishments that we're proud of
We successfully built a multi-service backend running locally via Cloudflare tunnels, integrating live iMessage read receipts and typing indicators with NVIDIA NIM models. We also passed 115+ automated worker tests to ensure our guardrails held strong.
What we learned
We learned that AI alone isn't enough for safety-critical applications; combining Nemotron with deterministic SpacetimeDB reducers and community data is the only way to ensure reliability.
What's next for Kavach
Securing a dedicated Photon line for native family group chats, deploying to Maincloud, and expanding to Android and WhatsApp for broader accessibility.