Step 01
Step 1
Six modes inside Add: Type / Chat, Photo, File, Voice, SMS, Bank / AA.
Documents & KYCRetail + B2Bpilot· marketing-claim · ingestion-tiers
36-format ingest (Tier 1-5).
Add a CSV, parquet, OFX, MT940, bank-statement PDF, image, audio note, or video — one tab, six modes.
The problem
Why this exists.
Real Indian households have data in formats no single app accepts. A statement PDF here, an SMS forward there, a voice memo from the family CA.
We want one Add tab that takes any of it without asking the user to convert anything.
How it works · 3 steps
From input to outcome.
- Step 02
Step 2
Files pass through a sandboxed extractor stack — Tier 1 tabular, Tier 2 documents, Tier 2.5 statement PDFs, Tier 3 images, Tier 4 audio, Tier 5 video.
- Step 03
Step 3
Every extracted row lands in the same canonical schema so downstream features do not see the format.
Under the hood
The models and routes behind it.
Models called
API routes
Internal-only — no public route handler.
Compliance posture
- HEIC and WebP decode in a RLIMIT_AS/CPU-sandboxed subprocess; SVG rejected outright.
- PAN + 12-digit Aadhaar redaction on every transcript and frame-OCR row before parquet write.
- DPDP §12 — every ingested artefact and its extracted rows are erasable.
What it's not
We accept many formats but we do not promise to extract perfectly from any of them. Bad scans, broken audio, and unusual MT940 dialects still need a human-review fallback.
Related · features & stories