Discovery Systems · 01Week 5
How Instagram Knows: Recommendation Systems from First Principles
A feed turns small signals into a useful prediction.
insights(2) · reading index
Insights
Notes and case studies about AI systems, retrieval, infrastructure, and cost. Start with a topic, or follow a series.
Topic map
A compact way into the writing. The categories are separate shelves, but the ideas overlap.
Discovery Systems
Three product-first notes on how feeds, first-use experiences, and search help people find what matters.
Read note →
Discovery Systems · 02Week 5
The Cold Start Problem: How Apps Begin Recommending Before They Know You
Personalization starts with product choices before behavior exists.
Discovery Systems · 03Week 5
Scentum: Semantic Search as a Product, Not a Feature
Search becomes useful when people can describe a feeling, not just a keyword.
GenAI
Retrieval, reranking, learned indexes, and practical notes around production GenAI systems.
GenAI Aug 2026
Your Agent Is a Contractor, Not a Chatbot
A 40-step agent that is 99% reliable per step gets it right 67% of the time. The fix is not a better model. It is knowing which actions deserve a phone call — with LangGraph interrupts in Python and TypeScript.
GenAI Aug 2026
Your Contracts Folder Is Not a Search Problem
A JSON inventory answers aggregation questions with 9,810 tokens instead of 663,181 — a 68x context collapse. Two LangGraph tools in Python, one Pydantic schema, and the honest benchmark that argues against the design.
GenAI Oct 2026
I Swapped Embedding Models and Half My Search Results Changed
Switching from text-embedding-ada-002 to text-embedding-3-small on this site's 330 article sections kept only 47.6% of the top-5 results. The new model was better (recall@5 64.5% vs 56.6%), but forgetting to re-embed the index made search no better than random, with no error.
GenAI Sep 2026
What We Mean When We Say the Model Thinks Step by Step
Writing steps down gives a fixed model the depth that long chains of reasoning need, and thinking models now do it themselves. On a local test, a three-call chain around a thinking model cost 1.25x the tokens for the same answers.
GenAI Jul 2026 Week 5
How Instagram Knows: Recommendation Systems from First Principles
A product-first introduction to recommendation systems with the math made explicit: implicit feedback, matrix factorization, two-tower retrieval, ranking, and evaluation.
GenAI Jul 2026 Week 5
The Cold Start Problem: How Apps Begin Recommending Before They Know You
How products make their first useful recommendation through onboarding choices, defaults, context, exploration, and early feedback.
GenAI Jul 2026
PyTorch as a Problem-Solving Tool: Tensors, Loss, and Gradients
A small Python exercise for learning PyTorch as a problem-solving tool: model the problem, write tests, represent state as tensors, compute loss, and use gradients only after the invariant is clear.
GenAI Jul 2026
RAG From First Principles: Evidence, Instructions, and the Shape of Refusal
A first-principles RAG exercise in Python: use tests to define evidence, keep retrieved documents as data, reject instruction-like text, and refuse when support is weak.
GenAI Jul 2026
Before the Model Answers: Retrieval, Reranking, and the Kraska Thesis
How B-trees, LSM trees, and vector indexes are converging on the same underlying problem — and what that means for builders of GenAI retrieval systems.
Project Reviews
Project writeups and case studies with the constraints, tools, and evidence kept close to the work.
Project Reviews Aug 2026 Week 4
Scentum and the ScoreVector: Turning Taste Into ML Engineering
Scentum and the ScoreVector: turning subjective taste into structured ML engineering dimensions.
Project Reviews Jul 2026 Week 5
Scentum: Semantic Search as a Product, Not a Feature
How Scentum can make fragrance discovery work when people know the feeling they want before they know the words to search for.
Project Reviews undated Week 1
GenAI Chatbot: The Better Labyrinth Game
A practical AI engineering review of building a browser-based Next.js Web UI RAG assistant for real users: retrieval, latency, prompt control, failure modes, and expectations.
Case StudyRAG support assistant
Ecommerce AI Chatbot
A production support assistant for ecommerce: semantic product search, context-aware answers, and ticket deflection measurement.
Case Studysemantic search
Scentum
A semantic fragrance search product: natural language queries, vector embeddings, and scoring over scent descriptors.
Concepts
Engineering fundamentals, Python notes, containers, data structures, and measurement-first habits.
Concepts Sep 2026
Ten Thousand Rows Were Never the Data. The Weight Was.
4.87 million payment events, 1,600 weighted rows, the same logistic regression. A local and Spark experiment on keeping counts instead of expanding them.
Concepts Jun 2026 Week 4
AI-First UI: Designing for Pleasant Interactions in 2026
What 'pleasant' means when the thing behind your UI reasons in probabilities: visible intent, graceful uncertainty, and recoverable mistakes as the new UI primitives for 2026.
Concepts Jul 2026
The Digest Is the Container. The Tag Is Just a Rumor.
Containers before Kubernetes: why a container's real identity is its digest, not its tag — and what that means for reproducible ML training, on CPU and GPU.
Concepts Aug 2026
DNS Was Never Just an Address Book
Twelve of twenty-one well-known domains share a Cloudflare or AWS nameserver. Five live DNS queries, lamest to strangest, show the moment a lookup stopped answering "where" and started answering "who is asking."
Concepts Jul 2026
Unstructured Was Never the Right Word
Frontier models turning video, audio, sensor streams, and DNA sequences into structured JSON — the extraction problem finally has a general solution.
Concepts Jul 2026
The Reasoning Loop Is a Game Loop That Learned to Think
What decades of game engine event loops teach us about agent reasoning loops — six patterns, with conceptual code, for tool-calling agents.
Concepts Sep 2026
The Memory Hierarchy Doesn't Break Where You Think It Does
Sweeping input size from a thousand elements to sixty-four million to test whether cycles-per-element stays flat across L1/L2/L3 boundaries — and catching a self-manufactured memory wall along the way.
Concepts Sep 2026
Zero Milliseconds Is Not a Score
0ms on LeetCode is a stopwatch rounding error, not a score — the real optimization target is cycles per character, measured with perf after the algorithm is already O(n).
Concepts Aug 2026
The Neo Geo Was Not Faster. Neither Is Apple Silicon.
The Neo Geo won on a 1.6× clock advantage. Apple's M-series wins while clocking lower. One number explains both — and it decides which model fits your device.
Concepts Jul 2026
Thinking About the Background: Node Clouds, Juice, and Random Walks
Why danstativa.dev's background is a pointer-reactive node network instead of a static graph-paper pattern — direct manipulation, game 'juice,' and random walks, with the full working code.
Concepts Sep 2026
What We Mean When We Say a Generator Is Lazy
A lazy generator stops when nobody asks for the next value. In a streaming endpoint the one asking is our own loop, not the reader: a silent client let the model write 28,573 tokens before the socket pushed back.
Concepts Jun 2026
Python Type Hints: Looking Under the Hood
Python has not become statically typed. What changed is the tooling around it — type checkers, linters, and editors that reason about your code before it ever runs.
Concepts Sep 2026
From Stacks and Queues to Agent Memory: Old Containers in New AI Systems
Five data structures from year-one computer science — queue, stack, hash map, graph, heap — are what actually keeps an AI agent from failing in production, not the model.
Closing Essay undated Week 4
The Humble AI Engineer: Less Magic, More Measurement
The humble AI engineer: less magic, more measurement across AI engineering, ML engineering, Python engineering, retrieval design, expected value, and product purpose.
Concepts Mar 2026
From Stacks and Queues to Graph Algorithms
From Stacks and Queues to Graph Algorithms: how primitive containers — stacks, queues, sets, heaps — power the higher-order algorithms of computer science.
Concepts Jun 2026
Python Concurrency: Threads, Processes, and the Shape of Waiting
Python concurrency explained from the primitive upward: threads for overlapping waiting, processes for CPU work, and why concurrency is not the same as parallelism.
Concepts Apr 2026
Python Virtual Environments for AI Pipelines
Python virtual environments are a primitive that gives AI pipelines a clean dependency boundary, making experiments, training, and serving easier to isolate and rerun.
SEO/GEO
Search and generative-engine optimization notes, with a preference for measured changes over claims.
SEO/GEO Jul 2026
Page Speed Is On-Site SEO Because Crawlers Are Readers Too
A short on-site SEO argument from Common Crawl to Core Web Vitals: before ranking, a page has to be fetched, parsed, rendered, measured, and turned into evidence.
SEO/GEO Jul 2026
Delete Before You Compress: the Second Optimization Pass
The sequel to the on-site SEO workout: a second pass built on one hierarchy — does a byte need to exist, block, run, or weigh this much? Page JavaScript went from 113 KB to 1.6 KB and first paint from 0.6 to 0.4 seconds, mostly by deleting instead of compressing.
SEO/GEO Jul 2026
The On-Site SEO Workout: Training a Prerendered Site to a Perfect Score
A real case study of on-site SEO in 2026: prerendered HTML, a render-blocking CSS diet, right-sized AVIF images, and immutable caching — taking a live site from 69 to 97 performance without touching its design.
Economics
Expected value, routing, retrieval costs, and decision-making for GenAI systems.
Economics Sep 2026
The Assignment Is Not the Test. The Conversation After It Is.
A live coding round samples one hour of thinking under conditions the job never has. A take-home assignment followed by a design review sees the work and the reasons behind it, and the unpaid hours are a commitment test that runs both ways.
Economics Jul 2026 Week 1
The EV+ of GenAI: Why AI Engineering Is Also Game Theory
Why AI engineering is also game theory: model routing, retrieval, tool calls, human escalation, and stopping rules as repeated strategic decisions under uncertainty.
Economics Jun 2026
The EV+ of GenAI, Part 3: Agentic Cost Tuning
Tune agentic AI costs with EV-positive reasoning budgets, gated tool calls, context diets, stop-loss rules, and production cost traces.
Economics Jun 2026
Measuring the EV+ of GenAI: Real Code, Actual Numbers
Run the EV formula with real Python code: three GenAI strategies, Hugging Face models on CPU, actual EV numbers — Strategy A: −$0.052, RAG: +$0.230, Router: +$0.230.
Economics May 2026
Measuring the EV+ of Agentic AI & GenAI
Measure the EV of every Agentic AI and GenAI inference call. EV formula, RAG vs agentic cost breakdown, optimal model routing, and equilibrium play for production systems.