SanskaraAI-agent

Inspiration Indian weddings are legendary for their grandeur, but also for their chaos. The average Indian wedding involves 15+ vendor categories, hundreds of guests, and intricate cultural rituals that vary by community. We saw families stressed, overspending, and struggling to coordinate disconnected service providers. We wanted to build more than a directory; we wanted a planner—an intelligent agent that understands need, culture, and logistics.

What it does Sanskara AI is an autonomous agent that acts as your personal Wedding Project Manager.

Context-Aware Planning: It knows your budget, guest count, and cultural preferences ("South Indian", "Vedic", etc.). Proactive Vendor Discovery: Instead of making you search, it proactively initiates searches using the Google Places API to find rated vendors near you. Deep Research: A dedicated "Search Specialist" sub-agent uses Gemini to verifying vendor reputations, pricing trends, and venue constraints. Action-Oriented: It doesn't just chat. It can generate payment links, schedule appointments, and update your real-time dashboard. Visual Intelligence: Users can upload design inspirations, and the agent (via Gemini Vision) analyzes them to suggest matching decor vendors. How we built it We built Sanskara AI using the Google Agent Development Kit (ADK) and Gemini 2.5 Flash Lite as the cognitive core.

Orchestrator Pattern: We used a main "Planner Agent" that delegates research tasks to a "Google Search Specialist" sub-agent. State Injection: To optimize the context window and accuracy, we developed a

inject_dashboard_state mechanism. Instead of the agent asking "What is your budget?", the system injects the live Firestore state (Budget, Location, Secured Vendors) directly into the system prompt before every turn. Grounding: We integrated the Google Places API to fetch real-world vendor data (Phone, Rating, Images), ensuring the agent never hallucinates a fake venue. Full Stack: The backend is FastAPI (Python), serving a React frontend. State is persisted in Google Cloud Firestore. Challenges we ran into Context Management: Early versions lost track of the budget or user preferences. We solved this by moving state out of the chat history and into a structured "System Injection" layer. JSON Reliability: Getting the LLM to consistently return valid JSON for UI components (buttons, vendor cards) was tough. We used Pydantic models and a strict output schema to enforce structure. Hallucinations: The agent initially invented vendor prices. We fixed this by forcing a "Directory First, Search Second" protocol, where it prioritizes real data from our verified directory or Google Maps. Resilience: When precise data isn't available, the agent uses heuristic models for estimation, such as: $$ Total Cost \approx Venue Rental + (Guest Count \times 5000) $$ Accomplishments that we're proud of True Agentic Behavior: We moved beyond "chatbot" territory. Sanskara AI proactively acts—scheduling, paying, and researching—without constant user prompting. Latency Optimization: Achieved near-human response times by optimizing the ADK pipeline and using Gemini 2.5 Flash Lite. Seamless Integration: Successfully connected a Python/FastAPI backend with Google Places, Firestore, and a React frontend in a unified, state-aware system. Contextual Resilience: The "State Injection" system proved robust, maintaining conversation context across long planning sessions without token bloat. What we learned We learned that Goal-Based Agents > Chatbots. Users don't want to chat; they want things done. By shifting our prompt engineering from "You are a helpful assistant" to "You are a Project Manager graded on execution," we saw a massive jump in utility.

What's next for SanskaraAI-agent Gemini Live Voice: enabling full 2-way voice conversations for on-the-go planning while driving. Vendor Portal: Building a side for vendors to respond to the agent's inquiries directly. iOS Native App: Porting the mobile web experience to a native iOS app for better notification handling. Multi-Modal Styles: Letting users upload a Pinterest board and having the agent auto-generate a matching vendor list using Gemini 2.5 Vision.

Built With

• fastapi
• google-places-api-tools:-vite
• javascript/typescript-ai-models:-google-gemini-2.5-flash-lite-frameworks:-google-agent-development-kit-(adk)
• languages:-python
• pydantic
• react
• tailwindcss-cloud-services:-google-cloud-firestore