It is what makes the rest of the deck believable — and it is exactly what the product
itself argues.
308 msEnd-to-end checkRTX 4050 laptop GPU, 4 s window.
64Calls at once0 dropped. Where testing stopped, not where it breaks.
0Bytes of audio storedNot to disk, not to a log, not to a temp file.
₹0Build cost21 dependencies, all permissive, licence-audited in the build.
Feasibility on five axes
TechnicalPretrained models only — no training, no fine-tuning, no dataset licensing. ~582 MB resident on a ₹60k-class laptop GPU.
OperationalRisk profiles, prescribed actions and human review fit an existing fraud-operations workflow.
Economic₹0 total cost. No cloud contract, no per-minute fee, no API key.
ScalabilityOne process, one CUDA context, bounded queue. 64 calls measured; capacity reported as a floor, never a ceiling.
PrivacyMemory-only audio, encrypted voiceprints, deployable inside an institution's own perimeter.
DeploymentFully offline after install — no internet, no account, nothing to provision.
Challenge → engineering response
| Challenge | What it changed in the architecture |
|---|
| Detectors degrade on real-world audio | The precondition gate exists because of this. The system degrades to a refusal, never to a confident wrong answer. |
| A detector may be unreliable on audio it was never tested on | A validation gate withholds the signal and returns NO_VALIDATED_SIGNAL rather than reporting a score we do not believe. |
| Signals can disagree | A disagreement beyond the threshold is reported as SIGNAL_DISAGREEMENT, not averaged into a middle number. |
| Phones block call capture at OS level | Measured on a real device rather than assumed; the enterprise path attaches to infrastructure the institution already owns (SIPREC or a carrier media stream). |
| No budget for a telephony account | Implemented against the documented protocol and tested against a local mock that speaks the same format. |
| GPU memory is finite | One process and one CUDA context with a bounded queue — extra processes would multiply memory and serialise anyway. |
| A model file could be silently swapped | Every weight file is pinned by SHA-256 and verified at load; a mismatch aborts startup. |
| Fusion weights are not calibrated | Stated openly as a rule of thumb. Calibration with proper cost reporting is named as the next version, not implied as done. |
!
Compliance position. The DPDP Rules were notified on 14 November 2025. A
voice embedding that identifies a person is personal data, not anonymised features —
so voiceprints are encrypted at rest, never returned by any API, and destroyed
cryptographically on erasure, with every creation and deletion audited.
