AI that respects zero-knowledge.
Lead scoring, decay, smart cadence, call summarizer, inventory match — all run client-side against decrypted tenant data. We don’t collect prompts. We don’t train on you. We don’t even see the inputs.
Shipped in sf_p310 + sf_p314. 14-day free trial, no card.
Lead scoring — explainable, not magic
Each lead gets a 0–100 score with the top reasons that drove it. No black-box vendor API. Built on a distilled model that runs client-side, with the score recomputed live as the lead's activity evolves.
- Top-3 reasons per score, in plain English
- Per-tenant calibration after 100 closed deals
- Bias audit: source, language, locality
- Manager override + retraining signal
Decay — leads cool, the score reflects it
A hot lead at noon may be cold by evening. The decay engine applies a half-life per source and surfaces "needs attention" before the lead falls off. No cron, no surprises.
- Per-source half-life
- Cooling notification at half-life threshold
- Auto-archive at zero
- Reactivation signal on inbound activity
Smart cadence — choose the next move
After a missed call, the cadence recommender picks the next channel: WhatsApp template, voice retry, email nudge. Trained on closed-deal patterns from your own tenant, not aggregated.
- Per-lead, per-stage recommendation
- Channel suggestion with confidence
- Skip-if-replied built in
- A/B test cadence variants
Call summarizer — 30-second close, 6 languages
After a call, you get a 3-line summary, next-best-action, and a candidate WhatsApp follow-up. Whisper-class STT runs in your browser. The transcript is encrypted before any write.
- STT on-device (English, Hindi, Marathi, Tamil, Telugu, Kannada)
- Summary + intent + sentiment
- Drafted follow-up WhatsApp text
- TRAI-compliant recording prerequisite
Inventory match — show three real options, fast
A buyer wants "2BHK in Powai under 1.8 cr, east-facing, parking". The matcher returns the top three units across your inventory that fit, ranked by buyer fit, with explanations.
- Constraint parser handles vague Indian phrasing
- Top-3 ranked by configurable weights
- Vector + rules hybrid retrieval
- Per-builder inventory respect (multi-builder brokers)
No vendor LLM — by design
No OpenAI. No Anthropic. No Google. Every model is open-weights, quantized, and shipped in our bundle. Your prompts never leave your device. Your data never trains a foundation model.
- Open weights: Phi-3, Whisper, MiniLM, distilled BERT
- WebGPU + ONNX Runtime Web inference
- No telemetry of prompt content
- Falls back gracefully on older browsers
Inference runs entirely on the tenant device using open-weight models we ship in the bundle. No prompt, transcript, or embedding leaves the device. We collect aggregate latency + error counts to keep models healthy — never input content. This is a hard contract, not a marketing line.
See it for yourself.
14 days, full features, no card. Your keys, your data, your CRM.