Shmuel Devpost Project Story

Inspiration

Many small businesses already run through WhatsApp. Stock checks, customer details, bank transfers, receipts, and delivery arrangements often live across chats and the owner's memory.

Shmuel started from a simple question: what would it take to give these businesses an AI operator without letting the AI invent stock, prices, payment details, or promises?

What it does

Shmuel turns a plain-language description of a business into a supervised AI sales setup.

Business owners can:

• Generate and review operating rules
• Add verified inventory, prices, images, and stock levels
• Approve bank-transfer details
• Test customer conversations in a Playground
• Capture orders and payment receipts
• Connect WhatsApp through Linked Devices
• Review, approve, reject, or take over AI replies

The AI answers from approved business data and escalates when it cannot safely answer.

How we built it

Shmuel is a Next.js application deployed on Vercel, with Turso providing tenant-scoped storage for accounts, inventory, orders, conversations, and jobs.

Groq serves OpenAI's gpt-oss-120b and gpt-oss-20b models for setup generation, customer replies, and grounding review. Llama 4 Scout handles image understanding.

WhatsApp support uses NanoClaw and Baileys through an isolated persistent runtime. The Vercel control plane communicates with this runtime using signed commands through a self-registering ngrok tunnel.

Codex and GPT-5.6 were used throughout the engineering process: tracing failures across the UI, database, runtime, Docker, WhatsApp, and production deployment; implementing features; reviewing architecture; and verifying fixes.

Challenges we ran into

The hardest challenge was connecting a serverless application to a persistent WhatsApp runtime safely.

We dealt with rotating ngrok URLs, QR pairing failures, stale connection records, native Node module incompatibilities, container runtime failures, delivery acknowledgements, and cases where the dashboard claimed a message was sent before it reached the phone.

Grounding the AI was another major challenge. A good prompt was not enough. Inventory, payment details, conversation memory, and order state needed to become explicit sources of truth shared by both the Playground and live WhatsApp.

Accomplishments that we're proud of

We are proud that Shmuel is more than a chatbot demo.

It has a shared conversation engine for testing and live messaging, inventory-aware answers, persistent order memory, owner-controlled payment confirmation, isolated WhatsApp runtimes, and a supervised approval workflow.

We are also proud that we tested it with fresh accounts and real WhatsApp connections, found serious production failures, and fixed them instead of hiding them behind a polished demo.

What we learned

We learned that trustworthy business AI depends more on system design than clever prompting.

The model needs approved data, clear authority boundaries, durable state, and an honest way to say "I don't know." Live integrations also need to be tested at every boundary: the browser, database, runtime, provider, and receiving phone.

Working with Codex and GPT-5.6 also showed us the value of an iterative founder-AI workflow: test the real product, challenge the result, trace the failure, implement the fix, and verify it again.

What's next for ShmuelOps

Next, we want to move the linked-device runtime from the supervised Mac/ngrok pilot to durable Linux infrastructure, expand support for the official Meta WhatsApp Cloud API, and make provisioning safely self-service.

We also plan to improve media handling, background job processing, monitoring, fulfillment workflows, and guarded automation while keeping owners in control of payments, sensitive decisions, and customer escalations.

Built With

• baileys
• codex
• colima
• docker
• flutterwave
• gpt-5.6
• groq
• libsql
• llama-4-scout
• meta-whatsapp-cloud-api
• nanoclaw
• next.js-15
• ngrok
• node.js
• openai-gpt-oss-120b
• openai-gpt-oss-20b
• pnpm
• react-19
• resend
• turso
• typescript
• vercel
• whatsapp-linked-devices