Rohit Rajan

Problem

Our consumer products needed a personalized feed that stayed fast under real production traffic.

Hard decisions

1. Fan-out, asyncFeeds are built by an asynchronous fan-out pipeline instead of being assembled on each request.
2. Cache processed feeds in RedisServing pre-processed feeds from Redis is what holds response times under 50 ms in production.
3. Fix the read path, not just the symptomAfter the outage: paginated queries and database-level indexes instead of unbounded collection reads, then a sweep of the codebase for the same pattern.

Problem

A legacy Node.js monolith sat under three products and had to be decomposed to support a wider digital transformation, without taking any of them offline.

Hard decisions

1. Split by domain ownershipServices were carved out along domain lines, each owned by a team.
2. Tenant-by-tenant cutover with reconciliation gatesNo big-bang switch. Each tenant moved only after reconciliation checks passed.
3. Legacy stays the source of truthThe legacy path stayed live as the source of truth until the new path passed two full business cycles. I wrote the migration playbook the team followed.

Problem

Production compute ran entirely on AWS (EC2, S3). I own the deployment, monitoring and scaling story for a consumer product with real user traffic, so cost and headroom were mine to fix.

Hard decisions

1. Hybrid, not a full exitCompute moved to Hetzner bare metal while AWS stayed part of the setup.
2. Nginx in front of everythingA single Nginx layer fronts both sides of the hybrid.

Problem

A RAG pipeline over an internal corpus. Handing back whole documents isn't precise enough, so it returns the specific passage with its surrounding context, and it has to respect time.

Retrieval quality is graded by evaluation scripts against a held-out question set, not by feel.

Hard decisions

1. Embeddings as a separate FastAPI async workerThe Gemini 768-dimensional embedding service batches and retries independently from ingest, so upstream timeouts and rate limits never block uploads.
2. FAISS to QdrantMoved production vector search to Qdrant so metadata filtering happens inside HNSW traversal instead of as post-filtering.
3. Measure before trustingAn evaluation harness scores retrieval strategies against a held-out question set.

Problem

Tracer-Cloud/opensre is a LangGraph-based framework for building AI SRE agents. Agents investigating incidents need to read production logs, and teams running self-hosted Elasticsearch had no log source for it.

This integration lets downstream agents query production logs during incident investigations.

Hard decisions

1. Ship it as a LangChain toolThe integration is wrapped as a LangChain tool so agents in the framework can call it directly.
2. A small httpx clientFive methods against Elasticsearch, kept narrow and fully covered by 31 unit tests.
3. Tests before featuresA separate test-only PR took the auth and AWS modules from 0% to 100% coverage with 34 unit tests, making future contributions safer to review.

Problem

Commitments get made in WhatsApp chats and then lost. This captures them where they happen and puts a countdown on your Mac's menu bar.

TickDown itself is an open-source macOS menu bar app that counts down your day: seconds, minutes and hours left today, this week, this month and this year.

Hard decisions

1. Read the WhatsApp Web DOMThe extension captures commitments straight from the page you're already using.
2. One shared Convex backendThe extension and the macOS app sync through the same Convex backend.
3. Rust FFI into SwiftA Rust FFI bridge connects into the native Swift/SwiftUI app.

Problem

Aviation enthusiasts and flight-sim pilots want to see aircraft and weather on the same map. Flightradar24 keeps weather behind a subscription.

The extension adds four live layers you toggle from a popup, with adjustable opacity and data timestamps.

Hard decisions

1. Zero dependencies, no build stepPlain JavaScript on Manifest V3, about 1,300 lines in total. Load unpacked and it runs.
2. Isolated contexts with a narrow bridgepage.js draws on the map, a bridge relays messages, and the service worker does the network work. Page requests are limited to named data types, so the page can never make the extension fetch an arbitrary URL.
3. Privacy firstOnly the storage permission, talks only to MSN's weather servers, collects nothing.
4. Cache in the service workerThe worker fetches, caches and decodes weather data, and caches map metadata for 3 minutes.