How Sachet Works — Architecture & Methodology

1. Core Principle: "Warn, Never Block"

Sachet is engineered as a defensive co-pilot for the citizen, not an authoritarian telecom filter.

Calls are never forcibly dropped or interrupted. Instead, as the dialogue unfolds, Sachet provides transparent, contextual warnings in simple conversational Hindi (e.g., "Digital arrest scam: police/CBI ka naam aur paise maangna ek hi script hai"). This empowers ordinary citizens and senior citizens to identify deception themselves and hang up with confidence.

2. Four-Stage Detection Pipeline

Rather than brittle keyword blacklists (which scammers effortlessly evade by rephrasing), Sachet evaluates the temporal escalation of the call through four integrated stages. While overall conversational risk scoring is driven by our fine-tuned MuRIL transformer, the human-readable warning explanations (reason) are mapped via curated regex signal rules to guarantee clear, reliable, and auditable guidance.

Stage 1: Text-Only Ingestion

Speech is transcribed via browser speech APIs or server streaming; raw audio is never stored on disk and session transcripts are deleted after the call.

Stage 2: Fine-Tuned MuRIL

Uses google/muril-base-cased fine-tuned on code-mixed Indian dialects to capture subtle emotional coercion and conversational arcs.

Stage 3: 20-Signal Psychological Engine

Extracts high-risk behavioral patterns: authority impersonation, arrest threats, secrecy demands, urgency, and OTP traps.

Stage 4: Consecutive Confirmation

Requires at least two consecutive high-risk turns to trigger severe alert banners, effectively eliminating transient false alarms.

3. Key Coercion Signals Detected

Scams adhere to recognizable social-engineering scripts. Sachet's engine tracks 20 distinct behavioral signals:

Signal Category	Common Scammer Phrase / Tactic	Risk Level
`authority_claim`	"Main CBI Cyber Cell se Inspector Rathore bol raha hoon"	High
`arrest_legal_threat`	"Aapke naam par arrest warrant jaari hua hai, case confidential hai"	Critical
`secrecy_isolation`	"Kisi ko mat bataiye, room mein akele rahiye aur video call chalu rakhiye"	Critical
`urgency_deadline`	"Aaj raat 8 baje tak account block ho jayega, turant action lijiye"	High
`asks_otp_pin_cvv`	"Aapke number par 6-digit verification code aayega, batayein"	Severe
`asks_money_transfer`	"Verification ke liye saara paisa RBI security account mein transfer kijiye"	Severe
`unsolicited_link_app`	"Play Store se Quick Support Electricity bill update app download kijiye"	High
`tells_to_verify_safely`	"Bank kabhi OTP nahi maangta, official branch mein visit kijiye"	Safe (Negative)

4. Zero-Audio Privacy Architecture

Because phone calls are deeply sensitive personal interactions, Sachet is designed from day one with strict privacy boundaries:

Text-Only Ingestion: The Sachet backend API never receives, logs, or stores raw audio.
RAM-Only State: Active call turns and rolling risk scores are stored exclusively in temporary volatile memory (RAM). Nothing is written to a permanent disk or database.
30-Minute Automatic TTL: All inactive sessions are permanently purged after 30 minutes.
Immediate Hang-Up De-allocation: Calling /v1/session/end immediately deletes all session memory.