Step 01
Step 1
Every chat turn passes through a middleware stack that classifies for distress, gambling, abuse, and debt spiral.
SafetyRetail
Crisis-aware safety guardrails.
Distress, gambling, and financial-abuse detection — built in, not bolted on.
The problem
Why this exists.
An advisor that keeps optimising returns while the user is being financially abused is participating in the harm.
Distress, problem-gambling patterns, and intimate-partner financial coercion are detectable in transactions and chat — and we have a fiduciary duty to notice.
How it works · 3 steps
From input to outcome.
- Step 02
Step 2
When a signal triggers, the response routes through crisis-fallback rather than the usual advisor.
- Step 03
Step 3
Pattern detectors are honest — every match logs an audit row that the user can review.
Under the hood
The models and routes behind it.
Models called
API routes
Internal-only — no public route handler.
Compliance posture
- DPDP §6 — purpose of safety capture is disclosed; opt-out drops you back to plain-advisor mode with the same chat history.
- Detection is advisory — Mukodi never blocks a vendor or freezes an account on its own.
- Every fired guardrail surfaces under audit-trail.
Interactive demo · mocked at the edge
See it work.
Try the Courtroom (mock)
Loading the demo…
What it's not
Guardrails reduce the risk of harmful answers — they do not eliminate it. The Courtroom + HallucinationCheck pipeline is a second line, not a guarantee.
Related · features & stories