Simor

Data Infrastructure for Production AI

Practical writing on AI data engineering, feature stores, and the infrastructure choices that determine whether AI systems work in production.

Performance Engineering for GenAI Inference: Batching, Caching & Quantisation
Performance Engineering for GenAI Inference: Batching, Caching & Quantisation
12 Dec, 2025 | 05 Mins read

A startup's GenAI application cost $0.42 per query at 15-second latency. At this rate, their Series A funding would last six months. The problem wasn't the model—it was unoptimized inference. Each req

Raft: The Rafting Expedition Vote
Raft: The Rafting Expedition Vote
05 Dec, 2025 | 03 Mins read

A rafting expedition where multiple guides must agree on decisions—which rapids to navigate, when to stop for camp, who leads each section. Without consensus the expedition fragments. Raft consensus w

Case Study: End-to-End RAG Platform for Customer Support
Case Study: End-to-End RAG Platform for Customer Support
05 Dec, 2025 | 05 Mins read

A SaaS company with 200 support agents and 10,000+ knowledge base articles had an 18-hour average response time and 23% first-contact resolution. Their largest enterprise client threatened to cancel a

Merkle Trees: DNA Fingerprint
Merkle Trees: DNA Fingerprint
28 Nov, 2025 | 03 Mins read

Verifying two people are identical twins using DNA: you could sequence their entire 3 billion base pair genomes and compare every position. Or use genetic fingerprinting: hash specific DNA regions int

Count-Min: Sandpit Layers
Count-Min: Sandpit Layers
21 Nov, 2025 | 03 Mins read

Thousands of children play at a beach, each leaving footprints. Tracking each child's visits individually becomes impossible at scale. Instead, imagine multiple shallow sandpits with different grid pa

AI in Regulated Industries: Compliance Patterns for Finance & Healthcare
AI in Regulated Industries: Compliance Patterns for Finance & Healthcare
21 Nov, 2025 | 04 Mins read

Deploying AI in regulated industries—banks, insurance, healthcare—requires more than technical excellence. A model that's a black box cannot satisfy regulatory requirements for explainability. Trainin

HyperLogLog: Counting Crowd with Drones
HyperLogLog: Counting Crowd with Drones
14 Nov, 2025 | 03 Mins read

Counting attendees at a massive festival: individual counting requires massive infrastructure for millions of attendees. Sampling small areas and extrapolating fails with uneven crowd distribution. Th

Benchmarking Vector Databases: Performance, Cost & Ecosystem
Benchmarking Vector Databases: Performance, Cost & Ecosystem
14 Nov, 2025 | 05 Mins read

A RAG application that works perfectly with toy datasets grinds to a halt at production scale. The vector database that benchmarked beautifully with 10K vectors performs terribly at 10M. The one that

Tries: The Word Ladder
Tries: The Word Ladder
07 Nov, 2025 | 03 Mins read

Word ladder games start with "CAT", change one letter to get "COT", then "DOT", then "DOG". Now imagine all possible words connected in a web where shared prefixes create natural pathways. That's a tr