Project

Redrob Console

Summary

Client-facing API dashboard for a routed Indic text stack: Sarvam’s open model and other open backends behind one Redrob completion surface. I owned onboarding, integration support, and per-customer cost tuning. Routed Indic text API

My role

Owned the client-facing dashboard contract, onboarding and integration support, and per-customer cost tuning against the routed Indic completion surface.

Purpose

Give Indian enterprise customers a dashboard and API keys to one routed Indic text endpoint, while we swap Sarvam and other open models behind Nest without breaking their clients.

Decisions

  • Ship a routed Indic stack, not a single closed vendor

    The product is an Indic-first text API. Sarvam’s open weights sit alongside other open models; routing chooses the backend while customers keep one Redrob completion contract and rrk_ keys.

  • Make the dashboard identical to production traffic

    Playground sessions and server keys share the same complete and SSE streaming paths. If the UI and the API diverge on auth or routing, every demo becomes a lie.

  • Treat streaming as a correctness boundary

    SSE and non-streaming responses had to enforce the same key scopes and routing rules. A stream that bypasses the router is a silent cost and quality bug.

Challenges

  • Routing cannot feel like a weaker model

    Cost savings only work if Indic quality stays near the strong path on the tasks those accounts care about. The product has to feel like one model, not a manual picker.

Impact

  • Self-serve Indic inference with open backends

    Accounts try the routed model in the console, mint keys, and integrate without binding to a single upstream vendor. Deployment work included tuning spend against each customer’s mix.

System

  • Redrob Console

    Indic routed

    API dashboard

    Client-facing API dashboard for a routed Indic text stack using Sarvam open weights and other open models behind one Redrob completion contract

    Console dashboard and Nest API for keys and completion; Sarvam and other open models stay behind routing

Screens

Customer sign-in

Entry point for Indian customer accounts. Shows this is an external inference product, not an internal ops screen.

Playground

Try the routed language model, including Indian-language assist, before creating production API keys. Meant to show the demo uses the same completion path customers will call later.

Translate

Dedicated translation into Indian languages such as Hindi. Shows language needs for that market are built into the product next to completion.

API keys and complete

API documentation for sign-in, customer API keys, and the complete/stream endpoints. This is what client engineers integrate against for the cost-saving routed model.