Solaris

Solaris is a sales-automation platform that does the annoying first-touch work for you. A lead comes in, from a web form or an email or a scribbled call note, and Solaris reads it, qualifies it, drafts a reply, routes it, and tells you what to do next. The whole intake-to-action loop is one structured model call.

What I built

• A multi-tenant lead-intake app (Next.js, TypeScript, React, Claude) that ingests inquiries from forms, email, and call notes and returns a full qualification, summary, drafted reply, routing decision, and next task at once.
• JSON-schema-constrained structured outputs to extract contact fields and classify lead quality (hot / warm / cold / spam) without the model inventing data. API keys stay server-side behind Next.js route handlers.
• A per-client configuration layer injected into the model’s system prompt, so qualification logic, routing, and reply tone change per business with zero code changes. That’s the part that makes “multi-tenant” mean something.
• A full sales pipeline on top: deal stages, value tracking, an activity timeline, automated stale-deal detection (it flags anything inactive 7+ days), and an LLM follow-up generator that drafts stage-aware, tone-matched outreach from deal history.
• Real-time live-call transcription with an on-screen assist layer that surfaces suggested responses to objections and questions while the rep is still on the call.

Why it matters

Solaris only works if you can trust the model’s output enough to act on it without a person checking first. That’s why everything runs through JSON schemas: a lead gets routed, a reply gets drafted, and a task gets created from validated fields, not from free text the model might have made up. Being live also means it meets input I never designed for: the misspelled, half-finished inquiries that look nothing like the happy path. Handling those without falling over is what separates this from a weekend build.