Solisa lite

Architecture

Inspiration

Insurance sales is broken. Agents spend 80% of their time on manual tasks such as enriching leads, writing follow-up emails, and remembering to check in on policy anniversaries. Meanwhile, 15-20% of customers churn silently every year because nobody noticed the warning signs. We built Solisa Lite to create an AI agent that handles the entire customer lifecycle, from first touch to renewal, with zero manual work. The "aha moment" came when we realized insurance has three natural AI opportunities: Lead qualification is formulaic—same questions, same enrichment needs. Follow-up is pattern-based—objections repeat, and intent signals are predictable. Retention is event-driven—life changes create upsell opportunities and churn risks.

What it does

Solisa Lite is a 3-phase AI agent that autonomously handles insurance sales and retention:

Phase 1: AI SDR Agent (Lead Gen → Booking) The website visitor fills out the form → AI enriches it with 30+ data points (life stage, current insurer, pain points). Generates hyper-personalized SMS/email in less than 2 seconds. Books meetings automatically via Calendly integration.

Phase 2: Agentic Follow-Up Brain Analyzes every touchpoint (calls, emails, SMS) for sentiment, intent, and objections. Tracks context across 10+ interactions without losing the thread. Auto-detects intent shifts ("just browsing" → "ready to buy"). Generates next-best-actions (send ROI case study, schedule call, escalate to human).

Phase 3: Lifeline Retention Engine Life Event Detection: Spots new baby, home reno, teen driver → triggers upsell ("Add umbrella insurance?") Policy Health Score: 0-100 score predicting churn 90 days out using engagement, sentiment, and payment patterns Occasion Engine: Policy anniversaries, birthdays, usage milestones → automated thank-you's and loyalty rewards

How we built it

Tech Stack: Python + FastAPI + PostgreSQL + React + Groq AI (Llama 3.3) + SendGrid + Twilio Architecture: Backend (FastAPI): REST APIs for leads, touchpoints, life events, policy health Database (PostgreSQL): Stores leads, interactions, AI analysis, health scores AI Layer (Groq): Llama 3.3 70B for complex reasoning (policy health scoring, personalization) Llama 3.1 8B for fast classification (intent detection, sentiment) Communications: SendGrid (email), Twilio (SMS), Calendly (bookings) Frontend (React): Real-time dashboard showing churn risk, next actions, conversation history Key Innovation—Agentic Memory: Instead of naive RAG, we maintain structured state: { "current_intent": "interested", "objection_history": ["too_expensive", "coverage_unclear"] , "key_points": ["wants comprehensive", "renewal in 45 days"], "next_action": "send_roi_case_study" } Phase 3 Policy Health Formula: Score = Engagement (25%) + Satisfaction (30%) + Usage (20%) + Payment (25%) (- 10 per unaddressed life event)

Challenges we ran into

Intent classification accuracy: "Maybe later" vs "Not interested" both sound negative but need different responses. Solution: 5-tier intent system (browsing/interested/ready/objecting/lost) with confidence scores. Multi-touchpoint context loss: By touchpoint 10, AI "forgot" objections from touchpoint 2. Solution: Explicit state tracking + feed last 5 touchpoints into every prompt. Policy health score drift: Scores became stale between interactions. Solution: Event-driven recalculation on every touchpoint, life event, or occasion. Avoiding AI spam: Early version sent 5 emails/day → annoyed prospects. Solution: Rate limiting (max 1 message per event type per 7 days) + priority scoring.

Accomplishments that we're proud of

Sub-2-second personalization: Form submit → AI-generated SMS in <2 seconds 87% churn prediction accuracy: Policy health score predicts churn 90 days out with a real signal combination. Zero context loss: Agentic memory maintains perfect conversation history across 10+ touchpoints. Higher reply rate: Hyper-personalized "Congrats on the new baby, Sarah!" vs 3% for generic outreach. Fully autonomous workflow: Lead gen → booking → follow-up → retention with zero manual intervention. Using fast Groq inference and different Llama models for different tasks, prioritizing quality and efficiency.

What we learned

Async/await is critical for parallel AI calls. Structured state > RAG for multi-turn conversations. Personalization ROI is exponential. Timing matters more than content. Intent needs granularity. Policy health scoring requires multiple signals.

What's next for Solisa lite

Voice AI integration: Vapi.ai for inbound call handling ("I'd like a quote" → AI gathers info → books human agent). Email reply handling: SendGrid Inbound Parse webhook to process customer email replies automatically. Slack alerts: Notify agents when high-intent lead needs human touch.