Writing

Blog

Focused essays on the engineering work behind AI products: real-time systems, LLM evaluation, hybrid search, recommendations, ranking loops, and the infrastructure that keeps everything safe to scale.

01 / Core essays

May 28, 20266m

Hybrid Search Without Hand-Waving

A practical blueprint for building hybrid search with Azure AI Search: lexical retrieval, vector search, semantic ranking, filters, scoring profiles, and relevance evaluation.

Read

May 26, 20265m

The Home ARR Lab

What a self-hosted Jellyfin, Jellyseerr, Sonarr, Radarr, indexer, and download-client stack teaches about queues, automation, observability, and operating real systems.

Read

May 24, 20265m

Flash: Talent Search at Scale

A product and architecture breakdown of Flash, a talent search and recommendation platform built around candidate signals, recruiter workflows, ranking, and scale.

Read

May 22, 20266m

Ranking Loops That Learn

Why search and recommendation quality improves through feedback loops, offline evals, judgment labels, exploration, and product instrumentation - not just a better model.

Read

Apr 12, 20265m

Building a Real-Time AI Interviewer: Voice, Turn-Taking, and Latency Budgets

How I think about real-time AI voice systems in production: audio streaming, speech detection, transcription, LLM turn-taking, and the latency budget that makes an agent feel alive.

Read

Apr 10, 20265m

Designing LLM Evaluation Systems: Rubrics, Consistency, and Guardrails

A practical blueprint for evaluating interview answers with LLMs: structured rubrics, multi-pass scoring, prompt versioning, bias checks, and guardrails that make the system safer.

Read

Apr 8, 20265m

Semantic Caching for AI Products: Cutting LLM Cost Without Breaking Quality

How to design a semantic cache for LLM-heavy systems using exact cache keys, embeddings, pgvector, Redis, thresholds, and quality checks.

Read

Apr 5, 20265m

Operating a Multi-Tenant AI SaaS: Isolation, Rate Limits, and Product Safety

The backend patterns I care about in AI SaaS: tenant resolution, data isolation, usage limits, billing boundaries, and infrastructure decisions that keep AI features safe to scale.

Read

02 / 2026

The ALTER TABLE That Killed My Deploy

How a single schema migration took down production for 4 minutes, and the expand-contract pattern that fixed it.

10m

Streaming LLM Responses in Production: SSE, Backpressure, and Per-Tenant Token Budgets in Next.js

How we built real-time AI feedback streaming for HyrecruitAI interviews — handling SSE reconnection, backpressure, and per-tenant cost attribution.