Family
Memory extraction pipeline
dormantApache 2.0multimodalinternal
Scribe.
Per-turn auto-capture pipeline. Pre-classifier abstains 90 percent of the time; for the remaining 10 percent the extractor proposes 5 factual MemoryTypes bound to a strict JSON schema.
Foundation · IndicBERTv2 (Apache 2.0) + Qwen 3-MoE (Apache 2.0)
Marketing claim
Scribe Service captures per-user memory with DPDP-strict 4-flag consent + minor short-circuit + cost cap.
What it does
The honest description.
Per-turn auto-capture pipeline. Pre-classifier abstains 90 percent of the time; for the remaining 10 percent the extractor proposes 5 factual MemoryTypes bound to a strict JSON schema.
Architecture
How it works.
Build per-user, consent-gated memory of factual mentions (income, dependents, named accounts, life events) to give downstream models context. SHADOW mode only — captures land in an audit table, not in user memory.
Training data & drift
What it learned from, and how we watch it.
Source
Synthetic multilingual eval corpus (HI-Roman / HI-Devanagari / TA / BN) at ≥250 turns/lang. No real user transcripts.
Drift baseline · current quarter
Drift baseline not yet documented for this model.
What it's NOT
- Ships disabled; `SCRIBE_ENABLED=false` default. No user memory is mutated until the operator flips the flag.
- DPDP-strict 4-flag consent + minor short-circuit + per-user daily cost cap — any of these veto capture.
- Pre-classifier is LightGBM-shaped (ONNX) with an abstention zone — operators see honest 'no decision' bands, not silent passes.
- Model-inferred fields are excluded from credit-risk features by a boot-time invariant.
Who calls this model
Across the house and the API.
Internal routes
No public routes documented yet.
Featured in
/features/your-memory16-char nonce injection-defense wrap
Per-record consent filter before any persistence
HMAC-SHA256 pseudonymized audit ledger
Multilingual factual extractor (5 MemoryTypes, JSON-schema bound)
Version history · immutable once published
Every release, traceable.
Data Mukodi holds about you is subject to DPDP §12 erasure (use /consent in the app), and any concern about this model can be raised via [email protected] under DPDP §13 grievance, or via /security.