logo

Outsurance: Your Privacy-First AI Insurance Agent

Inspiration

Millions of Indians remain uninsured or locked into plans that do not suit their actual health profile. Purchasing insurance today requires visiting multiple insurer portals and filling out repetitive forms with no personalized guidance. The core gap isn't a lack of products, but an intelligent matching layer: a system that ingests a person's real health data, processes it, and returns a ranked shortlist with clear reasoning. We wanted to build a platform that doesn't just list plans, but actively protects users from hidden exclusions and waiting periods, ensuring radical transparency.

What it does

Outsurance is a privacy-first, edge-deployed smart insurance recommendation platform. Users complete a guided multi-step intake form and optionally upload a clinical lab report. Using Local AI, we extract key metrics (HbA1c, BP, BMI) without the document ever leaving the user's local network.

The core of the platform is a 3-stage ML pipeline (XGBoost risk classification, 6-factor suitability scoring, and Cosine similarity KNN ranking) that evaluates 154 real Indian health plans.

The Agentic Layer: We integrated an omnipresent AI Agent powered by Jac to elevate the platform from a simple matching tool to a fully autonomous advisor. The Jac agent utilizes:

Memory: It retains the context of the user's specific 10D health vector and risk profile across the session.
Tool Use: Users can ask natural language questions (e.g., "What happens if I have a 5-day ICU stay?"), and the Jac agent autonomously triggers our FastAPI backend's Stress Test Emergency Simulator to calculate out-of-pocket costs and room-rent penalties.
Multi-Step Reasoning: It compares plans side-by-side, evaluating hidden conditions like co-payments and day-one exclusions, and generates plain-English explanations for why a plan is recommended or flagged with a warning.

How we built it

Agentic AI Engine: We used Jac (Jaseci) to orchestrate real agent behavior, replacing standard API wrappers with an autonomous agent capable of planning and executing multi-step queries (like live ML re-assessments and plan comparisons). Local Gemma (Edge SLM) was integrated for secure data extraction and natural language generation.
Frontend: A responsive, minimalist UI built with Next.js (App Router), TypeScript, and Tailwind CSS v4 to ensure a premium, high-end user experience.
Backend & ML: FastAPI + Uvicorn serving a Python-heavy ML stack (XGBoost for risk tiering, trained on 100,000+ records).
Database & Auth: Supabase (PostgreSQL) with strict Row-Level Security (RLS) and JWT authentication to guarantee user data privacy.
Deployment: Designed as an Edge Deployment capable of running on hardware like a Raspberry Pi, making it accessible in remote hospital waiting rooms without internet dependencies.

Challenges we ran into

True Agentic Behavior vs. Chat Wrappers: The biggest challenge was moving beyond a simple LLM chat interface. Implementing Jac required us to deeply integrate tool-calling capabilities so the agent could actually run our custom Stress Test Simulator and interact with the XGBoost pipeline, rather than just returning static text.
Privacy-First Edge Architecture: Ensuring absolute data privacy meant raw medical documents could never hit our cloud servers. Balancing the performance of local SLMs (Gemma) on edge hardware with the complex reasoning required by the Jac agent took significant optimization.
Scoring Complexity: Translating abstract insurance terms (like waiting periods and age gates) into a quantifiable 6-factor suitability mathematical model (Budget Fit, Condition Match, Risk Alignment, etc.) was a rigorous data engineering challenge.

Accomplishments that we're proud of

Agentic Orchestration: Successfully utilizing Jac to demonstrate real tool use and conversational memory, allowing users to run mock medical emergencies via natural language.
High-Accuracy ML: Achieving 87.1% accuracy on our XGBoost classifier for health risk tiering.
Edge Portability: Engineering the entire 3-stage recommendation engine and agent to function efficiently on edge hardware, ensuring remote accessibility.
Seamless UX: Translating a highly complex, data-heavy backend into a clean, "anti-gravity" style minimalist interface that makes buying insurance feel effortless.

What we learned

AI-Native Tooling is a Paradigm Shift: Using Jac taught us that building AI-native applications is fundamentally different from traditional full-stack development. Giving the agent the autonomy to plan and execute tools creates a significantly more dynamic user experience than rigid, pre-defined user flows.
Transparency Builds Trust: In InsurTech, users care more about why a plan is recommended and what its warnings are (e.g., a 4-year wait for diabetes cover) than just the price.

What's next for Outsurance

Expanding the Jac Agent's Toolset: We plan to integrate live hospital billing APIs into the agent's tool registry, allowing the Stress Test Simulator to pull real-time, hospital-specific pricing for highly localized out-of-pocket estimations.
Broader Catalogue: Expanding our dataset beyond the current 154 health plans to include term life and family floater specifics at a national scale.
B2B Integration: Pitching the edge-deployed hardware solution to offline clinics and local hospitals in Bengaluru to digitize their patient advisory process. Outsurance_Devpost_Writeup.md Displaying Outsurance_Devpost_Writeup.md.