Briskk — WhatsApp Store-Staff Copilot for Retail

Vertical AI copilot for omnichannel retail. Drives staff productivity and revenue by handling product discovery, live stock checks, and instant cart creation and follow-ups —on WhatsApp!! Pilot-live.

Comment

💡 Inspiration

Floor staff already sell on WhatsApt, but they juggle POS/ERP/Shopify and screenshots. Answers are slow, variant details get missed, and WhatsApp revenue isn’t attributed. We wanted a copilot that actually sells—fast, reliable, and tracked to orders.

✨ What it does

  • Compresses find → stock → price → compare/similar → cart → follow-up into one WhatsApp flow

  • Inventory check with multi-step clarifiers (store → variant)

  • Price inquiry (exact variant)

  • Add to cart with shareable, tracked checkout link (orders/chat attribution)

  • Lead capture & reminders (soft-launch)

🏗️ How we built it

  • Router → Plan → Execute pipeline; Redis for conversation/state (24h TTL)

  • Claude Code (~70%) for planning clarifiers and selecting intent-specific button labels

  • Postgres + pgvector for retrieval/similarity with price/stock validation

  • Interakt (WhatsApp BSP) + Shopify deep links (guest checkout) with UTM + order notes for attribution

  • Prometheus metrics, p95 latency target ≤3.5s; Dockerized, tenant-isolated config

🚧 Challenges we ran into

  • Variant & facility ambiguity → fixed with clarifier gating (ask store, then variant)

  • Low-latency on mobile networks → cached embeddings, deterministic API paths for stock/price/cart

  • Reliable tool use → minimal prompts, strict validators, clear fallbacks + escalation logs

  • Multi-tenant safety → tenant IDs at every hop; per-tenant greetings/buttons

🏆 Accomplishments we’re proud of

  • Pilot-live with Furnishka ( Bengaluru)

  • Five production intents end-to-end (discovery, inventory, price, add-to-cart, similar)

  • End-to-end attribution: link opens → carts/chat → orders/chat (72-hr)

  • Intent-specific button labels reducing clarifier loops and errors

🧠 What we learned

  • Clarify first, then answer beats LLM guessing for store/variant/pincode.

  • Tiny UX choices (button text, card layout) move completion rates.

  • Attribution plumbing turns a nice demo into measurable ROI.

🚀 What’s next for Briskk — WhatsApp Store-Staff Copilot for Retail

  • FAQ/Policy, pincode/ETA, offer/quote, and reminders

  • Public Shopify app; customer-facing agent reusing the same infra

  • 3–4 week pilot readout: p95 latency, % resolved w/o escalation, link opens → carts/chat → orders/chat, stock-out save rate, staff minutes saved

Demo link - https://drive.google.com/file/d/1iCLkHZWQumln0DmL2JkEhjkG6VVu-e27/view?usp=sharing

Built With

  • ai
  • claude
  • elixir
  • grafana
  • llm
  • multi-tenant
  • pgvector
  • phoenix
  • postgresql
  • prometheus
  • python
  • rag
  • redis-(conversation/state;-24h-ttl)-messaging
  • shopify
  • whatsapp
Share this project:

Updates