Rakshasakhi: one Gemma 4 model that sees and hears a scene, for women’s-safety triage in city control rooms.

Most video-AI pipelines turn speech into text first, losing fear in a voice, a scream, or the moment a shout coincides with someone blocking a path. Rakshasakhi gives Gemma 4 E4B the frames and the raw audio of the same moments, interleaved in time, so it judges the scene the way a person standing there would.

How it works

• Upload any clip (iPhone HEVC, WhatsApp, portrait, misnamed files). It is decoded in memory and cut into 4-second slices: frames + that slice’s audio.
• Instant alerts (<2 s, no AI): sudden loud sound or sudden movement are flagged at once, so an officer can act immediately.
• Gemma 4 verdict: category (harassment, following, surrounded, distress…), severity, the cause, the time span of the incident, and what was seen and heard in every slice, with speech quoted in the original language (Telugu/Hindi/English).
• Report in English, Telugu or Hindi. A human officer confirms every dispatch. Nothing is automated against a person.

Privacy & accountability by design (built for India’s DPDP Act)

• Runs entirely on the laptop: footage never leaves the machine.
• Officers see only face-blurred frames; blurring fails closed. Unblurred frames exist only in memory for the model, never stored.
• Evidence encrypted at rest; keyed, tamper-evident audit log of every view, playback and decision.
• Unreviewed alerts are never auto-deleted (flagged overdue); dismissed incidents and their audio are purged.

Honest limits: on a 24 GB MacBook the AI verdict takes ~1–2 min per 30 s clip (seconds on a GPU server). The small E4B model struggles with dark, chaotic footage. Accuracy has not yet been measured on a labelled test set; scripts/evaluate.py --compare measures sight+sound vs sight-only.