🍔🧠 The Personalization Stack DoorDash Built (Serves 100M+ Users)

PLUS: Uniqlo Decoded 🚨, Agentic patterns⚡, Be the idiot mindset 👨‍💻

Alexandre Zajac

Jul 13, 2026

Happy Monday! ☀️

Welcome to the 176 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

Separating signal from noise in coding evaluations matters
Meerkat: an experiment in global consensus building
AI code reviewer trust at DoorDash
Understanding is the new bottleneck in software
Some new agentic patterns emerging now

🗞️ Tech and AI Trends

New method keeps kids safe from illegal AI content
Grok 4.5 launches with major improvements
Apple sues OpenAI over trade secret theft

👨🏻‍💻 Coding Tip

React Server Components with Suspense streamline data-fetching, reducing time-to-interactive significantly

Time-to-digest: 5 minutes

Building a unified consumer memory for personalization at scale 🛒

DoorDash built a unified memory platform that extracts semantic understanding from behavioral signals and serves it to every personalization surface—from traditional ML models to LLM-driven experiences. Instead of treating user data as just a feature engineering problem, they treat it as a semantic extraction problem.

The challenge: Traditional engagement-based models can’t capture the why behind purchases. An embedding knows users buy kale and almond milk, but it can’t express “plant-forward health-conscious” in a way that generalizes across categories or LLM systems can reason about.

Implementation highlights:

Three-layer memory architecture: Long-term memory (durable preferences), in-session context (real-time intent), and explicit context (user-stated preferences) operate on different timescales
LLM-synthesized memory blocks: Use LLMs to distill behavioral patterns into semantic narratives—”Strong affinity for organic produce; price-conscious on packaged goods”—that both ML and language models consume
Versioned manifests for decoupling: Separate memory generation from encoding via manifests. Deploy new component versions to 10% of users, revert if quality drops. Reconstruction of any consumer’s memory at any historical date becomes trivial
Asymmetric dense embeddings: Map consumer memory and item descriptions into shared semantic space using block-specific retrieval instructions. Memory blocks act as high-level query expansion pulling items matching semantics users haven’t purchased
Consumer context graph: Build heterogeneous graph with consumers, brands, taxonomies, and semantic concepts as nodes. Keywords extracted from memories form semantic bridges enabling multi-hop reasoning—”prefers X → implies Y → connects to untried categories”

Results and learnings:

Semantic matching primitive unlocked: Memory blocks work for ML systems too, not just LLMs—many personalization problems are fundamentally about aligning consumer intent with catalog semantics
Multiple encodings beat single representation: Dense embeddings capture semantic similarity; graphs capture relational structure. Together they capture more signal than either alone
Versioning prevents cascading chaos: Full lineage (model ID, prompt hash, response hash, timestamp) means unexpected downstream behavior becomes traceable. You can reconstruct historical states and A/B test across generations

DoorDash’s move from “users click X, Y, Z” to “health-conscious, organic-preferring, premium-seeking” shows that scale isn’t incompatible with semantic understanding. Their layered architecture proves you can decouple how you generate memory from how you encode it—each improving independently on different timelines.

Introducing Meerkat- an experiment in global consensus

Cloudflare Research is building a global consensus service called Meerkat that uses a new consensus algorithm called QuePaxa. We plan to use Meerkat to build a strongly consistent, fault-tolerant key-value store, and other applications.

How we learned to trust our AI code reviewer at DoorDash

Agentic test processes, LLM benchmarks, and other notes on agentic coding from Galapagos Island

Deep notes from Dan on using AI effectively

Separating signal from noise in coding evaluations

A new analysis from OpenAI reveals issues in SWE-Bench Pro, a popular coding benchmark, raising concerns about reliability and accuracy in evaluating AI models.

ARTICLE (engineers go brrr)
The software engineering war

ARTICLE (shirt code chaos)
Decoding the obfuscated bash script on a Uniqlo t-shirt

ESSENTIAL (yeet yourself into jobs)
Advice to new grads: take any job you can get in the industry

ESSENTIAL (brains are the new problem)
Understanding is the new bottleneck

ARTICLE (robots go think-think)
Some new agentic patterns

ARTICLE (dumb but wise)
Be the Idiot

Want to reach 200,000+ engineers?

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

WORK WITH US

🤖 Chinese AI Models Surge in U.S. as OpenAI and Anthropic Costs Skyrocket (4 min)

Brief: Chinese AI companies like DeepSeek and Z.ai are rapidly gaining market share among U.S. businesses, with their models now accounting for up to 46% of tokens used on developer platforms, driven by significantly lower costs (60-90% cheaper) while maintaining near-frontier performance as American labs raise prices on their most advanced systems.

🚀 Former GitHub CEO Launches Entire, an AI-Optimized Git Hosting Rival (4 min)

Brief: Thomas Dohmke, ex-GitHub CEO, debuts Entire.io, a decentralized Git hosting network built to handle AI agent traffic that’s overwhelming GitHub’s infrastructure, offering agent auditing, code mirroring, and superior performance (2.1M pushes/hour vs. Cursor Origin’s 81K).

🤖 Apple Exploring Ways to Run Much Larger AI Models Directly on iPhones (2 min)

Brief: Apple is in talks with PrismML about running larger AI models on iPhones, with the startup successfully shrinking Alibaba’s 27-billion-parameter Qwen model to fit on iPhone 17 Pro—bigger than Apple’s current 20-billion-parameter on-device model—which could enable more Apple Intelligence features to run locally and reduce reliance on cloud servers.

🤖 SpaceXAI Releases Grok 4.5, Elon’s Answer to Anthropic’s Opus (2 min)

Brief: SpaceXAI launches Grok 4.5, an Opus-class AI model with twice the token efficiency of competitors and significantly lower costs ($2/$6 per million tokens vs. Opus’s $5/$25), positioning itself as a faster, cheaper alternative just as OpenAI prepares to release GPT 5.6 this week.

🚀 OpenAI Launches GPT-5.6 Sol, Terra, and Luna Across Apps and API (3 min)

Brief: OpenAI rolls out GPT-5.6 with three tiers—Sol (flagship), Terra (affordable), and Luna (fastest)—featuring multi-agent coordination, improved artifact generation for work outputs, and tiered pricing starting at $1 input per million tokens, now available globally across ChatGPT, ChatGPT Work, Codex, and the API.

🎙️ OpenAI Upgrades ChatGPT Voice Mode with New GPT-Live Model (3 min)

Brief: OpenAI launches GPT-Live, a new voice mode model that delegates complex tasks like web search and deeper reasoning to GPT-5.5 while maintaining conversation flow, finally replacing the outdated GPT-4o era model that limited ChatGPT’s usefulness as a brainstorming partner.

This week’s tip:

React Server Components with Suspense boundaries enable server-side rendering of expensive data-fetching components while streaming HTML and deferring client hydration. Use Suspense to wrap server components that fetch data and fallback UI shows while the Promise resolves server-side, reducing time-to-interactive.

Wen?

Large data-driven feeds (e-commerce, news): fetching 100+ items server-side and streaming HTML chunks avoids waterfall requests and improves CLS and LCP compared to client-side fetch-then-render.
Personalized pages with authenticated data: server components can safely access database/session context directly without exposing secrets through API routes or client bundles.
Mixed static+dynamic layouts: Suspense fallbacks allow fast static-shell rendering (header, nav) while deferred RSC chunks load personalized content below-the-fold asynchronously.