hash.ring · 6/6 nodes upclick a node to take it down
shipping backends for fintech and B2B SaaS, since 2018
8years
the workflow engine I designed at Neuron7, now shared infrastructure for 3 services
0→1
API response times at Rattle, after a multi-layer caching redesign
−65%latency
on Slice, where I built the KYC-document and credit-bureau services
10M+users
Designed for the bad day.
Every long-running workflow fails halfway through eventually. Four rules decide where it lands.
01
Decide the failure up front
Every step declares its policy when it's designed, whether that's compensate, degrade, retry later or stop.
02
Retries replay, they don’t repeat
Every side effect carries an idempotency key, so a retry returns the first result instead of doing the work twice.
03
Match the answer to the error
A timeout and a 422 deserve different responses. Transient errors get a retry; final ones get a decision.
04
Undo ships with the step
Compensation is part of each step's contract and unwinds in reverse order, nested workflows included.
Not every failure deserves a rollback.
A toy saga engine, running in your browser. Run it, then break it four ways.
order.fulfilmentsaga · 5 steps · 1 child workflow
ready0.0s
Break it
Press Run, then break it. Each step answers failure the way it declared.
order placed
compensateAuthorise paymentwaiting
compensateShip the orderwaiting
child workflow
Create labelwaiting
done
causenone yet
waiting for a run
What the toy leaves out
A compensation that itself fails. The run parks and a human gets a task.
Idempotency keys on the undo calls too, so a retried void or cancel can’t fire twice.
An outbox, so a notification leaves only after the state it describes commits.
Versioning for workflows that are already in flight.
Fairness between tenants sharing one engine.
Written for this page. General patterns only; no Neuron7 code or design.
Eight years in production.
Two B2B SaaS platforms, consumer fintech at 10M+ users, vernacular content and enterprise CSR, newest first.
Own the multi-tenant workflow stack end to end, from the custom execution engine (shared infrastructure, about 1,000 runs a day across 3 services) to Customer Workflows, the customer-facing product with its visual Workflow Builder.
Designed the engine from scratch, with recursive (nested) workflows, rollback and compensation, and developer-facing handler interfaces built from composable step primitives (Assignment, Notification, Integration for ERP pull and push, and AI Enrichment).
Built the User Tasks Service (human approval and task assignment) from the ground up, and re-architected the legacy Notifications and Integrations services into reliable step handlers.
Lead a team of 3 across architecture, delivery and code review, and shipped the Dev Kit documentation other engineers build against.
AI-powered service intelligence · B2B SaaSHybrid, Bengaluruneuron7.ai ↗
Architected Node.js / NestJS services with multi-layer caching, cutting API response times 65% and page-load times 50%.
Built Project Board, a Salesforce data-visualisation product, from scratch. Owned the backend API design, coordinated the React / Next.js integration, and shipped from MVP to production on customer feedback.
Stood up Docker microservices on AWS ECS (Nginx, ALB), built the GitHub Actions CI/CD, and defined auto-scaling that held through sustained production load.
Built the Document Management Service (encrypted KYC storage and retrieval with retention policies) and the Credit Reporting Service (durable bureau-data ingestion, storage and report generation), two platform services used across downstream teams.
Rebuilt the Jenkins deployment pipelines for reliability and configurability, reaching zero-downtime deployments.
Revamped the onboarding backend with A/B experimentation infrastructure, lifting engagement 25% and cutting drop-off 18%.
Built Node.js backend services and the supporting React UI, serving 100+ enterprise clients.
Enterprise CSR and volunteering platformBengalurugoodera.com ↗
Earlier
ACM ICPC India Regionals, Kolkata · 2016
CodeChef peak rating 1970, 41 rated contests
Head of Enigma, the college programming club · 2017–18
B.Tech, Computer Science, ABES Engineering College · 2014–18
Previously.After hours · A TV and anime companion for iPhone
Your shows, right where you left them.
Add a show once and its seasons, films and extras stay together, with progress for each season. Today says what to watch next, Schedule lays out the week, and Recap catches you up on what aired since your last visit.
Under the hood
AniList files every season, film and OVA as its own entry, so one show can span dozens. A breadth-first walk over the relation graph settles the obvious links, an LLM referees only the ambiguous side stories, and if the model is down the graph’s answer ships instead. TMDB covers TV, and a Fastify and Postgres service on a Mac mini keeps both in sync.
SwiftUI
Fastify
PostgreSQL
Drizzle
AniList
TMDB
Personal project · iPhone · coming soon
I choose boring technology.
What I reach for first, and the job each one is trusted with.
PostgreSQL
The source of truth. Constraints, transactions and the right index solve more than a new database would.
Redis
Hot reads, locks, rate limits and sessions. Fast and disposable, never the system of record.
TypeScript
Where I’m fastest. Typed contracts at every service boundary.
NestJS
The framework I reach for on new Node services. Modules and guards keep a large codebase legible.
AWS
ECS, Lambda, S3 and CloudWatch, kept as simple as the product allows.
Also in production
Node.js
Kafka
SQS
Java
Spring Boot
Python
Go
Express
GraphQL
MySQL
MongoDB
Elasticsearch
Docker
Nginx
GitHub Actions
Jenkins
React
Next.js
Let’s compare notes.
Workflow engines, failure handling, platform teams. Always glad to talk shop, and I usually reply within a day.
SwiftUI
Fastify
PostgreSQL
Drizzle
AniList
Node.js
Kafka
Java
Spring Boot
Python
Go
Express
GraphQL
MySQL
MongoDB
Elasticsearch
Docker
Nginx
GitHub Actions
Jenkins
React
Next.js