_1751731246795-BygAaJJK.png){kind=logo}
For Startups
Playbooks, decision frameworks, and case studies written for startups.
Cache LLM Embeddings in Redis Without Stale Vector Drift
Caching OpenAI embeddings in Redis cuts cost, but naive keys leave stale vectors polluting your retrieval layer. Here's a content-addressed cache pattern that self-invalidates on every edit — with runnable code.
How GimBooks Scaled GST Filing to 3M Users Without a DBA
GimBooks handled normal load fine but crashed every GST deadline cycle. Here's the architectural call that fixed it — and why horizontal scaling alone would have made the problem worse.
Supabase vs. PlanetScale vs. Neon: Pick the Right Serverless DB
Most Supabase vs PlanetScale vs Neon comparisons benchmark cold starts on toy workloads. Here's how each one actually behaves under multi-tenant SaaS load — and which choice will force you to re-platform when your first enterprise customer lands.
Stream LLM Responses to the Browser Without Losing State
Streaming LLM output to the browser looks trivial until you need per-user state, accurate token billing, and safe recovery when the client drops. Here is a working pattern with FastAPI, SSE, and React that gets all three right.
How Vahak Onboarded 800K Trucks Without Burning the Database
Vahak's onboarding pipeline scaled from hundreds to hundreds of thousands of trucks without breaking live matching queries. The unlock wasn't sharding or caching — it was separating write and read paths at the data model level.
Rate-Limit LLM API Calls Across Workers Without a Queue
Your OpenAI feature works fine on one worker and dies on four. Here's how to build a distributed token-bucket limiter in Redis with an atomic Lua script — no queue, no broker, no rewrites.
Stream LLM Responses to a React Frontend Without Melting
Your ChatGPT-style feature stalls for 6 seconds before rendering a single token. Here is how to stream LLM responses to React properly — with auth, aborts, and partial JSON that does not double-render on flaky networks.
What Is Idempotency? Stop Charging Customers Twice
A double-charge after a mobile timeout is almost always a retry bug, not a payment gateway bug. Here's how idempotency keys work, and why the fix lives in your client — not your server.
Rate-Limit an LLM API Without Dropping User Requests
Watching 429s spike in Sentry every morning? The fix isn't smarter retries — it's a sliding-window token ledger that holds requests locally until your budget refills. Here's a runnable Python tutorial.
Pinecone vs. pgvector: Which Vector Store Fits Your AI App
Filtered vector search is the query pattern that breaks most head-to-head Pinecone vs pgvector benchmarks. Here's how to pick the right vector store for your AI app based on the dimensions that actually matter in production.
How GimBooks Served 3M Users Without Breaking GST Logic
A walkthrough of how the GimBooks accounting SaaS handled GST edge cases at scale by treating compliance as a state machine, not a calculation library. The lesson generalizes to any fintech whose rule logic works at 50K users but silently breaks at 500K.
What Is a BFF? Why Your Mobile App Deserves Its Own API
A shared API for web and mobile sounds efficient until your mobile team is making four round-trips to render one screen. Here's why the Backend for Frontend pattern is really about org structure, not network hops.