• 

  The Bridge Agent: When a customer scans a QR code in-store, the associate instantly receives a briefing with their preferences, reservations

• 

  "Mom" has voted for Option 1, "Dad" has approved the ₹5,000 budget with a thumbs-up emoji, and WEAVE (AI agent) is summarizing the consensus

• 

  Style DNA™: We don't just recommend products; we build a 512-dimensional vector of every user's fashion identity to predict fit and returns

• 

  Gemini-Native Discovery: Understanding "Hinglish" voice notes, processing celebrity photos, and grasping intent (weddings, urgency) better

• 

  Built entirely on Google Cloud, utilizing a 3-tier memory architecture (Hot, Semantic, Deep) to maintain context across millions of tokens.

• 

  WEAVE AI: The Nervous System for Commerce. Connecting WhatsApp, Web, and Store into a single, continuous conversation powered by Gemini 3.

WEAVE AI: The Operating System for Indian Commerce

---

Inspiration

The inspiration for WEAVE came from a realization of the ₹52,000 Crore Betrayal happening daily in Indian retail. We noticed that despite spending years buying from specific brands, customers are treated like strangers every time they initiate a new interaction.

We identified four specific "betrayals" that drive this disconnect:

1. The Stranger Treatment: Brands have zero memory of past context, leading to repetitive "How can I help?" loops.
2. The Language Wall: 240 million Tier 2/3 customers are locked out because apps don't speak regional languages or understand "Hinglish".
3. The Sizing Lottery: Customers order multiple sizes and return most of them because recommendations are generic, costing billions in reverse logistics.
4. The Solo Player Trap: Indian shopping is a family activity, yet every e-commerce app is built for a single user, ignoring the reality of family decision-making.

We realized that brands don't have bad products; they have amnesia. We built WEAVE to cure it.

---

What it does

WEAVE is India's first Gemini-native Commerce Intelligence Operating System. It is not a chatbot; it is a multi-agent nervous system that connects WhatsApp, the Web, and Physical Stores into a single, unbroken fabric.

The system relies on Thread Memory™, which unifies every interaction into a persistent history. WEAVE operates through a constellation of 7 specialized agents:

• Discovery Agent: Handles multimodal inputs (text, voice, image) to find products using intent rather than keywords.
• Style DNA™ Agent: Builds a mathematical representation of a user's fashion identity to predict fit and preference.
• Rescue Agent: Detects abandoned carts or failed payments and intervenes proactively to save the sale.
• Bridge Agent: Connects the digital and physical worlds. If a user browses online and walks into a store, the store associate's tablet is already updated with their preferences and reserved items.
• Family Sync Agent: Turns shopping into a multiplayer experience, allowing families to vote, comment, and pay for a shared cart via WhatsApp.
• Proactive Commerce Agent: Uses external triggers like weather or calendar events to initiate helpful sales conversations.
• Voice Commerce Agent: Enables an end-to-end shopping experience via voice notes for users less comfortable with typing.

---

How we built it

WEAVE is architected entirely around Google's Gemini 3 capabilities, utilizing a 3-tier memory structure hosted on Google Cloud.

The Brain: Gemini 3

We chose Gemini 3 because it is the only model capable of handling our requirements:

• 2M Token Context Window: Essential for remembering conversations from months ago.
• Native Multimodal & Multilingual: Crucial for understanding "Hinglish" voice notes and processing image inputs simultaneously.
• Function Calling: To execute real-world actions like checking inventory or sending UPI payment links.

The Architecture

We implemented a Thread Memory™ system with three layers of persistence:

1. Hot Memory (Redis Stack): Stores the active session, cart, and last 5 messages for $<5ms$ latency.
2. Semantic Memory (Vertex AI Vector Search): Stores the user's Style DNA and conversation embeddings.
3. Deep Memory (Firestore): A permanent log of purchases, family relationships, and life events.

We used Gemini 2.0 Flash for the orchestration layer to ensure sub-500ms responses, routing complex reasoning tasks to specific agents.

---

Challenges we ran into

• Handling "Hinglish" Nuances: A major challenge was accurately parsing mixed-language voice notes like "Bhai yeh wala thoda loose hai" (Brother, this one is a bit loose). We solved this by leveraging Gemini's native multilingual audio understanding rather than using a separate transcription layer, allowing the model to capture the intent and tone directly.
• The "Bridge" Latency: Syncing a user's online behavior to a physical store tablet the moment they scan a QR code required optimizing our data pipeline. We used Cloud Pub/Sub and Cloud Run to ensure the store associate receives the "Briefing" in real-time.
• Orchestrating 7 Agents: Managing state between the Family Sync agent (multi-user) and the Rescue agent (time-sensitive) was complex. We built a central Orchestration Layer using Gemini 2.0 Flash to route intents to the correct agent dynamically.

---

Accomplishments that we're proud of

• Style DNA Vectorization: We successfully mathematically modeled a user's style. By analyzing returns and purchases, we generate a $512$-dimensional vector that reduced return rates in our simulations by 37%.
• The "Phygital" Handshake: We are proud of the Bridge Agent. Seeing a store associate's tablet light up with "Priya prefers A-line cuts" the moment she walked in—based on her WhatsApp chat from yesterday—was a magical moment.
• ROI Potential: We calculated that for a large retailer, WEAVE could generate an ROI of 45,000% by recovering lost revenue from abandoned carts and returns.

---

What we learned

• Memory is a Relationship: We learned that without memory, AI is just a search engine. With memory, it becomes a relationship. The ability to recall a wedding mentioned 45 days ago changes the user's trust level entirely.
• Shopping is Multiplayer: Building the Family Sync Agent taught us that Indian commerce is fundamentally collaborative. Ignoring the "Mom approval" factor is why so many carts are abandoned.
• Voice is the Great Equalizer: By enabling full voice commerce, we realized we could unlock the digital market for millions of users who can speak their intent but struggle to navigate complex English UI.

---

What's next for Weave AI

• Scale to 100M Users: We have designed a cloud-native architecture using Google Cloud Load Balancers and Auto-scaling Cloud Run containers to handle national scale.
• Expanded Proactive Triggers: We plan to integrate deeper with Google Maps and Weather APIs to allow the Proactive Agent to suggest purchases based on hyper-local weather changes (e.g., suggesting raincoats when a storm is predicted in a specific city).
• Enterprise Rollout: We aim to pilot the SaaS model with mid-sized retailers, offering the "Growth" tier that includes 100K conversations and 50 store syncs.

Built With

• bigquery
• c/c++
• cloud-armor
• cloud-load-balancer
• cloud-pub/sub
• firestore
• go
• google-cloud-run
• google-gemini-2.0-flash
• google-gemini-2.0-pro
• google-maps
• google-weather-api
• next.js
• python
• razorpay
• redis-stack
• shiprocket
• shopify
• vertex-ai-vector-search
• whatsapp-business-api