Thamba

In the last 19 months, senior citizens in Pune lost nearly ₹100 crore to a single phone script: the "digital arrest." We read a heartbreaking report about a 72-year-old losing ₹1.05 crore over six days because a fake CBI officer on a video call forced them to stay on camera and cut ties with their family. The government has built verification tools (like the ED's QR summons check and CBI's ABHAY). But these tools assume you have a physical document, you're calm enough to solve a captcha, and you're sitting at a computer. The digital arrest scam is a call. We needed a tool that works mid-call, while the victim is isolated. Thamba ("stop" in Marathi) is that live smoke alarm. What it does:

Thamba operates directly in the browser to analyze scam calls and notices in real-time.Live Audio Check: Hold it near the phone. It transcribes the Marathi/Hindi audio, identifies the scam script, and speaks its verdict out loud.Photo Verification: It reads photos of arrest warrants or ED summons, checking for legal requirements (e.g., magistrate signatures, addressed to police, lack of QR codes).Escalation: It breaks the victim's isolation by sending a one-tap WhatsApp alert to their family.Smart Govt Routing: It extracts data like the 16-character CNR number and deep-links the user to the correct government portal to verify it themselves.

How we built it:

Because a panicking senior needs a verdict that cannot be wrong in the dangerous direction, we modeled the risk using strict boolean logic rather than relying purely on probabilistic AI:$$\text{Scam Alert} = (\text{Authority Claim}) \land (\text{Isolation Tactics}) \land (\text{Money Demand})$$Thamba works in three layers:Deterministic Rules Engine: A lightning-fast (<100ms) offline rules engine that knows the core scam script in three languages.Google Gemini: Acts as our AI layer to transcribe Marathi voice inputs, read notice photos, and catch novel phrasings (like "break all FDs").Routing Layer: Extracts the right variables and hands the user off to official verification bodies. 🚧 Challenges we ran into

AI Hallucinations: We couldn't afford Gemini hallucinating a "safe" verdict on a scam call. We built a strict guardrail: Gemini must quote the caller word-for-word, or its flag is dropped. It is also sandboxed as an "escalate-only" tool; it can never clear a warning triggered by the deterministic rules engine.No Public APIs: Government verification sites are protected by captchas and OTPs, so we couldn't verify cases automatically. We solved this by extracting the exact strings (like the eCourts CNR format) and deep-linking the user so they can paste the number effortlessly.

🧠 What we learned

We learned how to design AI as a constrained, defensive layer. By pairing a strict rules engine with Gemini's ability to read Devanagari script and parse forged images, we built a system that is instant, hallucination-proof, and deeply practical.🚀 What's next for ThambaWe want to move the entire engine into a WhatsApp bot. Instead of opening a web browser, a panicking senior can simply forward a scammer's voice note or photo warrant to the Thamba contact and get an instant verdict.