Manan Chawda

Writing

Field notes from
production systems.

Practical writeups on backend platforms, data infrastructure, and production AI. The filter is simple: lessons from systems I have shipped, rebuilt, or studied closely enough to explain.

Production AI

RAG, evaluation, agents, vector ingestion, and the infrastructure around model behavior.

Distributed systems

Kafka, streaming, idempotency, reconciliation, queues, consensus, and failure boundaries.

Operating at scale

Cloud migrations, observability, data freshness, cost, reliability, and team-facing trade-offs.

2026

  1. 01

    Run the RAG eval before you redesign the architecture

    A production RAG lesson: evaluation should tell you whether to add agents, not become the thing you bolt on after the graph already exists.

  2. 02

    Real-time vector ingestion is a data platform problem

    Fresh retrieval depends on unglamorous infrastructure: Kafka backpressure, idempotent upserts, embedding retries, DLQs, and freshness metrics.

  3. 03

    What batch-to-streaming migrations teach you about reliability

    Lessons from moving rewards and data workloads from scheduled batch jobs toward streaming systems with clearer ownership, latency, and failure boundaries.

  4. 04

    Multi-Agent RAG with LangGraph: when 4 agents beat 1

    Building a production retrieval-augmented system with router, retriever, verifier, and synthesizer agents — and what changed when we added RAGAS evaluation.