• Hungry Minds
  • Posts
  • 🍔🧠 How DoorDash Uses LLMs To Predict 10M+ Customer Grocery Lists

🍔🧠 How DoorDash Uses LLMs To Predict 10M+ Customer Grocery Lists

PLUS: Google Stack Breakdown 🛠️, S3's Trillion-File System 💾, Compression Algorithms ⚡

Author

Alex Zajac
October 6th, 2025 • Time-to-digest: 11 minutes

Today’s issue of Hungry Minds is brought to you by:

Happy Monday! ☀️

Welcome to the 224 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

🗞️ Tech and AI Trends

👨🏻‍💻 Coding Tip

  • Query patterns and indexing smartly

Time-to-digest: 5 minutes

Big thanks to our partners for keeping this newsletter free.

If you have a second, clicking the ad below helps us a ton—and who knows, you might find something you love. 💚

✅ No custom code
Visually built the entire workflow logic. No more fragile scripts to maintain.

✅ Pre-built trusted APIs
Connected instantly to PagerDuty, Datadog, and Slack from Postman's huge API library.

✅ Modular and adjustable
When a schema changes, update 1 node in the flow in minutes. No code deployment needed.

DoorDash transformed how they solve the cold-start problem for grocery recommendations by leveraging restaurant order history. Instead of recommending random items to new grocery customers, they built an LLM-powered system that analyzes what you order from restaurants to predict what you'll want from their grocery stores.

The challenge: Scale LLM inference to 200+ million users without burning through seven-figure costs per run while maintaining recommendation quality.

Implementation highlights:

  • Tag-based compression: Replace raw order data with structured dish/cuisine/dietary tags to dramatically reduce LLM context size

  • Offline tagset mapping: Pre-compute mappings from unique tag combinations to grocery taxonomies weekly instead of running individual LLM calls per user

  • RAG-powered retrieval: Use K-NN search to narrow 1000s of grocery categories down to ~200 relevant candidates before LLM inference

  • Hybrid scoring model: Combine recency and frequency signals with LLM relevance scores to personalize recommendations for each user

  • LLM-as-judge evaluation: Implement automated quality assessment using LLM judges with metrics like nDCG@3 and quadratic weighted kappa

Results and learnings:

  • Massive cost savings: Achieved ~10,000x cost reduction compared to naive per-user LLM approach

  • Statistical significance: Observed measurable improvements in order penetration for both convenience and grocery verticals

  • Quality at scale: Successfully maintained recommendation relevance while serving 200+ million active users

DoorDash proves that smart architecture beats brute force when scaling LLMs. Their approach shows how proper data compression and offline pre-computation can make expensive AI affordable while maintaining quality.

Who knew that your obsession with Korean BBQ could help you discover the perfect kimchi selection? Your delivery history has been judging your grocery game all along!

Postman AI Agent Builder

Build AI agents faster with Postman AI Agent Builder. Access verified APIs, the latest LLMs, and tools for seamless workflows.

API Versioning - A Deep Dive

#92: Understanding API Versioning Strategies (14 Minutes)

The secret architecture behind "username already taken"

When you try to sign up on a platform like Instagram and type in your username, the system almost instantly tells you whether it’s available or not. If it’s taken, it even suggests alternatives on the spot.

Taking a Look at Compression Algorithms

Dissecting various compression algorithms.

Vibe Coding Is The Only Future - Steve Yegge

I just finished an absolutely mind-bending 100+ minute conversation with Steve Yegge, and I need to tell you - this man just called 99% of programmers “dinosaurs using Stone Age tools.” Coming from the guy who accidentally leaked that legendary 3,700-word Google+ rant while drunk, joined Amazon as employee #250, and now claims he can write 12,000 lines of production code per day using AI...

Inside Google’s Engineering Culture: the Tech Stack (Part 2)

Deepdive into how Google works from SWEs’ perspectives at the Tech Giant: planet-scale infra, tech stack, internal tools, and more

My Most Important Learning in Software Engineering

The irony is that my most important learning in software engineering wasn’t really about software engineering at all.It’s a lesson I learned painstakingly throughout the early years of my career, absorbing it through countless missteps and moments of doubt. Yet it’s only been in recent years that I’ve truly started heeding my own learning and advice.

Companies Should Stop Obsessing Over AI Tools And Do This Instead

Good culture is by far a better productivity hack than a shiny new AI tool!

ESSENTIAL (code getting cheaper)
Programming deflation

ARTICLE (jack-of-all-trades-master-of-fun)
Generalists in 2025

ARTICLE (paper-eating-robots)
Advanced Document Processing using AI

ARTICLE (fuzzy-address-finder)
Address matching using a fault tolerant trie

ESSENTIAL (bug-squashing-101)
Debug like a PRO

Want to reach 190,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

Brief: OpenAI launches Instant Checkout feature allowing direct purchases through ChatGPT from Etsy and soon Shopify merchants, while open-sourcing its Agentic Commerce Protocol to potentially reshape online shopping dynamics against Google and Amazon's dominance.

Brief: OpenAI's new AI video app Sora hits 164,000 installs in two days and reaches #1 on the US App Store, surpassing ChatGPT and Google Gemini despite being invite-only and limited to US and Canada.

Brief: Apple releases its Foundation Models framework with iOS 26, enabling developers to create on-device AI features in apps across health, education and productivity, with free AI inference and offline privacy protection.

Brief: Anthropic launches Claude Sonnet 4.5, an AI model capable of 30-hour autonomous operations and advanced coding capabilities, representing a significant leap from their previous 7-hour record and positioning itself as a leader in AI agents and computer use.

Brief: Meta's CTO reveals plans to become the software backbone for humanoid robots through a major investment comparable to their AR projects, focusing on solving complex challenges like object manipulation rather than hardware manufacturing.

This week’s coding challenge:

Build Your Own Shell

Real-world proficiency projects designed for experienced engineers. Develop software craftsmanship by recreating popular devtools from scratch.

This week’s tip:

ClickHouse skip indexes dramatically improve query performance on large tables by eliminating data blocks before reading. Combine minmax, set, and bloom_filter indexes strategically based on query patterns.

Wen?

  • Time-series analytics: Skip entire date ranges when filtering by user cohorts or event types

  • Log aggregation systems: Rapidly filter multi-TB tables by specific error codes or trace IDs

  • E-commerce analytics: Efficiently query product interactions across millions of user sessions

As you grow older you will discover that you have two hands. One for helping yourself, the other for helping others.
Audrey Hepburn

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).