Wrap any client — svitch.wrap(openai.client). Context and tools carry across every model.
PII is detected and redacted before the prompt ever leaves your codebase. Eleven Indian entity types, on by default.
Each query is complexity-scored. Routine work is auto-routed to a cheaper model — up to 80% off.
Runs entirely in your browser — the same engine runs server-side, before any model sees a single token.
Drop Svitch in front of the SDK you already use. No prompt rewrites, no proxy config, no new mental model.
OpenAI · Anthropic · Gemini · Llama
+ any OpenAI-compatible endpoint
Detects Aadhaar, PAN, UPI, IFSC and seven more Indian entity types and strips them out before data reaches any model. Reversible, deterministic, logged.
Routine queries score low and route to cheap models; hard ones escalate. You set the ceiling, Svitch spends under it — cutting model cost up to 80%.
Every prompt, redaction, route and response is hashed into an append-only chain. Break a link and the whole trail flags. Export for any auditor.
An encrypted on-premise enclave runs open models inside your own network over a WireGuard tunnel. Zero egress — sensitive data stays put.
A Hyperledger consent chain records who agreed to what, when. Prove lawful basis for every data subject — ready for DPDP grievance officers.