Shah Tirth — Blog
Thoughts on software development, AI, and building things that matter.
Why Retrieval Quality Beats Model Quality (And Nobody Believes It Until They See It)
Everyone's first instinct when a RAG pipeline gives bad answers is to upgrade the LLM. Almost every time, the model was never the problem — the retrieved context was garbage. Here's why.
Designing Tool Schemas That Don't Confuse the Model
Function-calling contracts are the real API surface of your agent. Bad names, vague descriptions, and dead-end errors don't just slow the LLM down — they make it guess. Here's what I learned building
Write the Test Before the Code: TDD in NestJS with Jest and In-Memory MongoDB
Stop writing tests that justify your code. Write tests that define it — using NestJS, Supertest, Jest, and in-memory MongoDB to ship backend features with real confidence.
Introducing @void-snippets/angular: Signal-Native CRUD for Angular 19+
Type-safe, batteries-included data fetching for Angular with MaybeSignal reactivity and full optimistic stack support.
From 3.1s to 320ms: How I Cut Voice AI Response Latency by 90%
Reduce real-time voice AI latency from 3.1 seconds to 320ms under 50 concurrent calls — with Twilio Media Streams, streamed LLM output, and parallelized TTS.