Social Q

The Challenge

Generic AI writing tools return the same output regardless of social context. Users needed something that understood conversation nuance — connection level, timing, tone, intent — and adjusted its analysis accordingly, without inventing unsupported claims.

What We Built

• →Multi-stage AI workflow: deterministic routing, structured model output, analysis, interpretation, and send-strategy generation
• →Safety controls for ambiguity, unsupported relationship claims, message-direction errors, and hold / do-not-send recommendations
• →Screenshot-to-text extraction for analyzing conversation images
• →Voice input via Deepgram speech-to-text
• →Threaded conversation history with optimistic UI updates
• →Anonymous-first identity with authenticated account migration and ownership scoping
• →Usage metering and entitlement logic with atomic quota operations and billing webhook handling
• →Evaluation tooling covering safety-specific failures, hold behavior, clarification quality, and hallucinated patterns

What Makes It Different

• ✦Situation Intelligence router — skips clarification for clear inputs, adds depth for ambiguous ones
• ✦Dynamic intent theming — the entire UI accent color shifts based on the user's selected communication goal
• ✦Three input modalities in a single composer: voice, screenshot, and text
• ✦Optimistic thread submission with a pending overlay so responses feel instant on mobile

Architecture Highlights

• ·Multi-level routing pipeline: context classification → bouncer → analysis → response generation
• ·OpenAI structured outputs enforce consistent JSON schema across all analysis modes
• ·Anonymous-first session model upgrades to authenticated profile on demand
• ·Deterministic safety gates sit in front of generative stages so the model cannot invent unsupported facts unchecked