The problem ProofRoute AI solves

Most AI products force a bad tradeoff:

• If you send full prompts to cloud models, you risk exposing sensitive data.
• If you avoid cloud models, you lose access to high-quality AI help.
• If you use subscription-based plans, you often overpay when usage is low.
• If you choose models manually, you can waste budget or get weaker results.

AI Router solves this by combining privacy, pay-as-you-go billing, adaptive model selection, and cryptographic accountability.

What people can use it for

• Sensitive AI workflows (legal, finance, healthcare, internal ops) where raw text should not leave the device.
• Developer/copilot workflows where API keys, passwords, or wallet addresses can appear in prompts.
• Cost-controlled AI usage with x402 USDC micropayments (no mandatory subscription; pay per request).
• Smarter model usage via adaptive orchestration (tier + speed/quality preference + domain + capability constraints).
• Auditable AI operations with Groth16 proofs and on-chain verification on Base Sepolia.

Challenges I ran into

1. x402 settlement timing vs. usage logging
   A key hurdle was that x402 settlement data (tx hash) is available after the request handler returns.
   I solved this by recording usage first, then patching the latest record inside onAfterSettle, so each usage item can be bound to a real payment tx before proof generation.

2. Startup reliability with payment middleware
   x402 route configuration can fail asynchronously (facilitator/network issues), which can crash or destabilize startup.
   I handled deferred errors explicitly and configured middleware to avoid blocking startup when the facilitator is temporarily unavailable, with clear dev-mode fallback behavior.

3. PII false positives vs. usability
   Contextual PII detection can over-mask useful context.
   I tuned Presidio thresholds and added guardrails (for example, only mask LOCATION when it looks like a specific street address), then merged detections with regex results by severity to keep protection strong without destroying prompt utility.

4. Keeping routing resilient
   If server-side routing is unavailable, the app still needs to function.
   I added local fallback routing so users can continue working even when POST /route is unreachable.

Use of AI tools and agents

The system uses multiple AI components that cooperate in a pipeline:

Edge privacy/analysis layer (local)
Regex + Presidio detect sensitive entities.
Text is masked into placeholders before cloud calls.
A reversible mask map is retained for safe post-response restoration.

Adaptive orchestration layer
The app sends routing metadata (tier, speed/quality weight, domain, thinking/web-search needs, context length) to the gateway.
The gateway selects the most cost-effective provider/model for that request.

Cloud model execution layer
Selected providers (Anthropic/OpenAI/DeepSeek/Gemini) process only masked content.
Responses are returned and placeholders are restored locally for natural UX.