Beyond Chatbots: The Real-World AI Agent Behind MealOBox’s Operations

MealOBox offers flexible meal subscriptions delivered reliably from vetted local kitchens. Customers can choose their plans, set delivery slots, skip meals when needed and enjoy consistent, healthy food.

But as we scaled, new challenges appeared:

• Customer queries piling up on WhatsApp

• Kitchen coordination issues

• Inconsistent service quality.

That’s where AI entered the picture.

How I Got Into AI

My background is in Cloud and DevOps, where I learned that intelligent automation unlocks operational scale. I realized that AI could do the same for MealOBox - making our support, logistics, and kitchen quality faster, safer, and more personal.

Through my network, I came across the AI Agents Masterclass. I joined because I wanted something practical - a way to wire AI into our daily operations.

What stood out during the session was the concept of composable agents with explicit tool use - like integrating Sheets, vector databases, and memory patterns through n8n. The big mindset shift for me was to treat agents like ops teammates with SLAs and escalation rules. Human-in-the-loop isn’t a fallback; it’s a feature.

Meet Dhaniyaa & Boxie - Our Ops Copilot

Our first AI agents, Dhaniyaa and Boxie, now act as MealOBox’s Ops Copilot. They handle routine but high-volume tasks that once overwhelmed our support and kitchen teams:

• Auto-triaging WhatsApp and in-app queries (order status, skip meals, address corrections, invoices/refunds)

• Suggesting delivery ETAs using kitchen readiness and live traffic data

• Guiding kitchen partners on prep lists and payout FAQs

Customers, kitchen partners, and delivery partners all use these agents. The result? Fewer manual interventions, faster responses, and cleaner escalations.

With the information, learning and exposure at the Masterclass, we set out to rebuild our Agents - stateful, tool-using agents with retrieval, typed outputs, and clean human handoffs. Reliability jumped immediately.

Behind the Scenes: The Tech Stack

Our AI Ops Copilot runs on a stack that balances flexibility with reliability:

• Runtime: n8n (Agent node with streaming + short-term memory)

• Model & Embeddings: Google Gemini

• Knowledge Base: In-memory Vector Store fed via Google Drive watcher and ingestion flow

• Data Sources: Google Sheets tools for customer profiles, subscription data, and orders

• Apps & Backend: React Native, Node.js on AWS (EC2, S3/CloudFront), MongoDB, Redis

• Messaging & Payments: MSG91, FCM, Razorpay

• Logistics: Google Maps APIs for routing, geocoding, and ETA predictions

The flow is straightforward: incoming queries trigger the n8n Agent, which classifies intent and calls the right tools (Vector retrieval for FAQs, Sheets for live lookups). Responses stream back in real time, with context preserved for follow-ups.

We also implemented strict guardrails - tool-only answers for sensitive ops, schema validation, PII-safe logging, and auto-escalation on low confidence.

Some of our biggest “aha” moments:

• Multi-pass geocoding and quick-confirm UI reduced failed deliveries.

• Typed schemas eliminated hallucinations.

• Live transcripts and one-click “take over” built ops team trust.

Early Results

We’re currently at the MVP stage, piloting internally and with a few kitchens. The feedback has been excellent:

• Common FAQs now auto-resolve.

• Skip-meal and invoice flows are one-tap.

• Escalations come pre-contextualized.

We’re already seeing lower response times, fewer WhatsApp backlogs, and smoother human–AI handoffs. We’re now instrumenting formal metrics as we scale.

Advice for Other Startups

If you’re thinking of building AI agents, here’s my advice:

• Start narrow, win fast. Pick one high-volume, high-ROI flow and ship.

• Design for handoffs. Human-in-the-loop builds trust and protects SLAs.

• Constrain your agent. Use retrieval-first logic, tool-only actions, and typed I/O.

• Measure everything. Track resolution rates, deflections, and escalation quality.

• Iterate weekly. Review transcripts and failures - it’s the fastest way to improve.

What’s Next

Over the next 3–6 months, we’re rolling out to more vendors in Lucknow, expanding coverage, and adding:

• Language and voice support (Hindi/Hinglish, IVR handoffs)

• Managed vector storage (Qdrant/Redis/Pinecone)

• Rider shift planning and micro-routing

• Kitchen QA audits for A/B/C grading

• Continuous evaluation tied to ops SLAs

• Cost control policies for frequent intents and caching

AI isn’t replacing our operations - it’s amplifying our people.

Let’s Connect

If you’re building in food-tech, logistics, or operational AI, I’d love to exchange notes or pilot ideas together.

🌐 website: https://www.mealobox.in/
📧 devish2@mealobox.in
🔗LinkedIn: https://www.linkedin.com/in/devish2/