Features · Compliance Co-pilot

Three layers — engineered for citation, not just chat.

Where the question comes from. How it's answered. What you can verify after.

Layer 1

Inputs — meet your team where they already chat.

WhatsApp Business API channel

The default channel. Each seat is a phone number; numbers are added to your WhatsApp Business account on onboarding. Messages route through Meta's BSP layer. Read receipts, typing indicators, image attachments (for notices), and voice notes (transcribed) all supported.

User WhatsApp → Meta BSP → Sahayak inbox

Telegram + Web fallback

For teams in regions or contexts where WhatsApp isn't the right channel — finance teams in multinationals often prefer Telegram or a web widget. Identical capabilities, identical audit trail.

Telegram bot / web widget → router

Hindi / Hinglish / English

A language classifier runs first; the retrieval still happens over the English corpus, but answer generation flips to the requester's register. Tamil / Marathi / Gujarati are Enterprise add-ons; ask sales for the rollout calendar.

Lang-detect → en-corpus → reply-lang

Image OCR for notices

Photograph a GST notice; the copilot extracts the section, demand period, demand amount, due date, and surfaces a draft response. Image OCR is opt-in per tenant.

Photo → OCR → notice-parser → draft

Layer 2

Processing — engineered for the moments AI usually breaks.

Intent classifier across 18 regulators

Every question is routed to a regulator-specific retriever (CBIC, RBI, SEBI, MCA, …). Mis-routing is the leading cause of bad answers; the classifier is monitored.

Retrieval over re-indexed corpus

Each regulator's circulars + notifications + master directions are re-indexed weekly. Embeddings + keyword + section-number hybrid retrieval. Stale answers are caught at the citation-validity check.

Citation-validity gate

No answer escapes without at least one citation that links back to a paragraph in the corpus. If retrieval is empty, the answer becomes "I don't know — escalate."

Trajectory traces via Langfuse

Every inference is captured as a span tree: classifier → retriever → LLM call → validator. Admins replay traces in the Langfuse UI; flag bad ones for next-week's evals.

Prompt-injection defence

Layered input classifier + output validator. Caught attempts are logged with classification only — never with full content beyond a hash, per DPDP.

Escalation to human

"I don't know" routes to a human escalation queue: your own internal CA panel by default, our partnered CA firm as a backup (Enterprise add-on).

Layer 3

Outputs — citations first, summaries second.

Cited WhatsApp reply

The bottom of every answer carries a Sources block — circular numbers, dates, paragraph IDs. Tap a citation to receive a verbatim quote with the link to the source PDF.

Answer + Sources block + tap-for-verbatim

Admin trajectory replay

Growth + Enterprise admins get a Langfuse-backed trace UI. See what was retrieved, what was generated, where the citation came from. Tag bad traces; they enter our weekly eval set.

Trace tree → tag bad → eval pipeline

Conversation audit log

Append-only conversation history per seat. Exportable on demand. Retention is configurable per tier; 30d default on Starter, 90d on Growth, custom on Enterprise.

Append-only log → export CSV

Weekly digest email

Every Monday, admins get a digest: question volume, regulators most asked about, top flagged answers, suggested re-training subjects. Helps compliance leads spot the gaps in their team's knowledge.

Weekly digest → action items → team review

Comparison

Co-pilot vs ChatGPT / Google / your CA partner.

CapabilityGoogle searchChatGPTSahayak Co-pilot
Cites the source regulationSometimes (stale)HallucinatesAlways
India-specific corpusMixedGenericYes, re-indexed weekly
Lives in WhatsAppNoNoNative
Audit-logged conversationsNoNoYes, DPDP-compliant
Prompt-injection defenceN/ALimitedTwo-layer
Admin review of bad answersNoNoLangfuse trace tagging
Hindi / HinglishLimitedLimitedNative
"I don't know" honestyPretends to knowConfidently wrongEscalates to human

Bring three real questions to the demo.

We'll answer them live on WhatsApp, with citations. If we're wrong, we explain why.