SRH Engineering

The consultancy's records had outgrown the tooling around them: tens of thousands of clients, tens of thousands of projects, and staff whose first action every morning is a search. When search takes seconds, people stop using the system and go back to asking each other.

The brief was a platform that stayed responsive at that size and could be handed to a client team afterwards, which made predictable performance and a boring, well-typed codebase the two requirements that mattered.

Next.js with TypeScript on the front, Tailwind and Shadcn UI for the interface layer, RTK Query holding all server state. Behind it, MongoDB with indexes designed around the actual access patterns rather than the entity diagram, and a tokenized search index that answers across more than twenty million records in under a second.

Client

Next.js with TypeScript, Tailwind and Shadcn UI components.

Server state

RTK Query with explicit cache tags; no parallel local copies.

Data

MongoDB, with compound indexes shaped around real query paths.

Search

A tokenized index built for prefix and partial matching across 20M+ records.

Operations

EC2, PM2, Husky-gated commits, staged deploys.

Deployment sits on EC2 with PM2 for zero-downtime restarts, and Husky gates commits so the type and lint contract holds even under delivery pressure.

1. 01

   Tokenized search instead of regex lookups

   Naive substring matching over that many documents cannot stay fast. Records are tokenized on write into a purpose-built index, so reads are index hits rather than scans, and partial-name lookups behave the way staff expect.

   Trade-off Write paths got heavier and the index needs reconciliation when documents change shape.

2. 02

   Pagination and projection as defaults, not options

   Every list endpoint pages and returns only the fields the view renders. Nothing in the app is allowed to request an unbounded collection, which keeps the worst-case response predictable as the data grows.

   Trade-off Aggregate views need dedicated endpoints rather than filtering a full fetch on the client.

3. 03

   TypeScript strictness enforced in CI

   The platform was always going to be handed onward, so types and lint rules are gated at commit time by Husky and again in CI. The next developer inherits a codebase that tells them what a payload contains.

   Trade-off Slower onboarding for contributors used to loose JavaScript, and occasional friction on urgent fixes.

20M+

Records in the search index

<1s

Query response at that size

70,000+

Projects, across 60,000+ clients