Food & Agentic
tanhAI
Vague cravings, group orders and saved reels, turned into real orders across Swiggy food, Instamart and Dineout by Jaggu, a monkey who speaks 11+ languages.
- Timeline
- 2025 – now
- Status
- Active
- Stack
- Python, Swiggy MCP, Sarvam 105B, Sarvam 30B, Saaras, Agentic loops
- Problem
- Ordering food is a dozen taps across apps, and worse when all you have is a mood, a group that wants different things, or a reel you saved and never opened.
- Bet
- Make ordering a conversation. Say what you feel like, who it is for, or paste a reel, and let an agent work out the rest and do the tapping.
- Built
- An agent on Sarvam models that listens to reels, plans the order and fills a Swiggy cart through MCP, fronted by Jaggu.
- Result
- About 65% less time to order, and sessions that average 4.2 minutes.
Works today
- Ordering by chat or voice in 11+ languages
- Intent from a vague craving to a concrete order
- Group orders, sorted in one conversation
- Food, Instamart and Dineout through Swiggy’s MCP
- Reel to order, from the reel’s audio
- A mindful nudge before checkout
Limits
- Reels are read by audio only: a dish that is shown but never said is missed
- Orders go through Swiggy only
Status
Active, v1.2.0
A 2-minute read
Ordering food should be a conversation. More often than a dish it's a mood, a table of friends who all want something different, or a reel saved and never opened. Tell tanhAI what you feel like and who it's for, or paste the reel, and Jaggu, a monkey who speaks 11+ languages, works out what you mean and fills a Swiggy cart across food, Instamart and Dineout without the app ever opening, on Sarvam's models from intent to speech.
How it works
Say what you want by text or voice, in any of 11+ languages: a dish, a mood, an order for the whole group, or a pasted food reel.
Reels are handled by audio alone: the soundtrack is extracted and transcribed to pull out dishes, cuisines and restaurants.
Sarvam 30B parses the intent quickly; Sarvam 105B reasons over it and plans the order.
Swiggy’s MCP checks real-time stock and local prices and fills the cart across food, Instamart and Dineout, without opening the app.
Before checkout Jaggu may push back. Save a gym reel in the morning and reach for a burger at night, and he’ll mention it.
What I built
Jaggu, the multilingual monkey
The heart of the app: a playful monkey persona who speaks 11+ languages, so ordering feels like chatting with a friend rather than tapping through a delivery app.
Sarvam model stack
Sarvam 30B handles fast intent parsing, Sarvam 105B is the main reasoning model, and Saaras powers the speech layer that lets Jaggu listen and talk back.
Swiggy MCP integration
The agent connects directly to Swiggy across food delivery, Instamart and Dineout: real-time stock checks, local pricing and cart building without loading the app UI.
Craves reel pipeline
Paste a reel and tanhAI extracts and transcribes its audio to find dishes, cuisines and restaurants, then checks what is available nearby. No vision model; it is purely audio-driven.
Decisions
Listen to reels, don’t watch them
The pipeline reads a reel’s audio, which already names most dishes and places. Leaving out a vision model keeps it simpler and cheaper. It gives up dishes that are only shown and never said.
Two model sizes
The 30B model takes intent, where speed matters most. The 105B model is kept for the reasoning that actually needs it.
A character, not a form
Jaggu turns a checkout flow into a chat with a friend who speaks your language. The personality is the interface.
Nudge, never block
The mindful check points out a gap between what you saved and what you are ordering, then lets you decide.
Outcome
- ~65%
- Less time to order
- 4.2 min
- Average session
- 11+
- Languages Jaggu speaks
Now
Version 1.2.0
Added Swiggy Instamart support and tuned Jaggu's language selector to prevent unsolicited banana ordering sprees.
Stack
Python, Swiggy MCP, Sarvam 105B, Sarvam 30B, Saaras, Agentic loops
Keeper of this file
Jaggu
The monkey inside tanhAI. Speaks 11+ languages, most of them about bananas. Boop him and see.
← Poke it
Review
“Oye! I asked for one banana and Instamart sent twelve dozen. Rachit's intent parser is too powerful. I am swimming in bananas. Five stars, would press buttons again.”