AAvinash Thiyagarajan

Avinash Thiyagarajan · Bengaluru

I build backends from zero to one.

Five years, four companies, each one closer to the foundation. Today I am the backend architect and founding engineer behind BondScanner, a SEBI-regulated bond platform that trades on NSE. I lead its backend team, and I am re-architecting it alongside ULTRA by Tap. This is that story in five short parts.

5 years in software4 companies2 platforms being re-architectedAbout 5 minutes
Read the story How I got here
Scroll to begin
01

I got to start from zero

BondScanner is an online bond platform, regulated by SEBI, that places every trade on NSE. I am its founding engineer and backend architect, and I lead the backend team.

BondScanner lets anyone find, compare and buy listed bonds from a phone.

It is an ,Online Bond Platform Provider (OBPP)SEBI's category for platforms that sell listed bonds online. Each one has to be run by a registered stock broker. run by a SEBI-registered stock broker and NSE member. That makes its backend unusual: every order has to go through a stock exchange, settle through a ,Clearing corporationThe exchange's settlement arm. It stands between buyer and seller, so money and bonds change hands safely. and stand up to a regulator's rules.

BondScanner · backend

Avinash Thiyagarajan

  • ✓
    Backend architectdesigned the investment journey
  • ✓
    Founding engineerbuilt it from 0 → 1
  • ✓
    Backend leadruns the team

Live in productionIn progress: re-architecture

It is already carrying real money

BondScanner's own published figures, as of October 2026.

1.2 lakh+investors on the platform
₹125 Cr+in bonds transacted
4.7App Store rating, from 66 ratings
45%of investments come from tier 2 and 3 cities

Source: BondScanner's App Store listing and its July 2026 investor-data release. These are the company's figures, not my own audit.

The investment journey, end to end

This is the part I architected: how an investor's money becomes a bond in their .Demat accountAn account that holds securities such as bonds in electronic form, with a depository like NSDL or CDSL. The exchange and the regulator are in the loop the whole way. Pick a step.

One investment, five steps
BondScanner backendthe journey I architected Inventory Payments Trade Settlement Holdings Bond supplyunits · price Paymentpartners NSEstock exchange Clearingcorporation DepositoriesNSDL · CDSL BondScanner backend Inventory Payments Trade Settlement Holdings Bond supply Payment partners NSE Clearing corporation Depositories

A simplified view. Vendor names and internal service design are left out on purpose.

My part in it

Architect

I designed the systems money moves through

Payment integrations, trade placement on NSE, settlement and holdings: the whole investment journey, plus the inventory management system that decides what can be sold.

Founding engineer

I built it from zero to one

I laid the engineering foundation that the product and the rest of the team now build on.

Backend lead

I run the backend team

Hiring, mentoring, and making sure the work gets shipped.

BondScanner is a team's work. The apps investors use, and many of the features in them, were built by other engineers. This page covers the backend architecture I designed and the team I lead.

And now, the second architecture

The first architecture gets a product to market. I am now designing the next one, for two platforms at once.

Step 1 · Done

Start from zero

Lay the engineering foundation and build the backend team.

Step 2 · Live

Run it for real

1.2 lakh+ investors, and ₹125 Cr+ in bonds transacted through it.

Getting from zero to one was the first job. Getting from one to what comes next is the one I am doing now.

02

I also build the tools my team works with

Leading a team means making its day easier. These are the AI initiatives I started at work: a bot that debugs support tickets, one set of skills for every AI coding tool, and a fixed path from idea to shipped feature.

A bot that answers the support channel

Support posts a problem. The bot works out whether it is a real bug, investigates it, and drafts a reply with the root cause, the fix and a query to verify it.

1
Classify first

Half the channel is routine ops requests. It skips those.

2
Investigate cheaply

A model 300× cheaper reads the code, logs and data.

3
Review with a stronger model

A stronger model checks the draft before anyone sees it.

4
Dry-run by default

Until switched on, drafts come to me, not the channel.

5
Never post twice

A reference on each reply stops a crash repeating it.

# tech-supportillustrative
SU
Support11:42

An investor paid, but the order still shows as pending. Can someone take a look?

B
Support botAPP11:46

Looked into it. Here is what I found:

  • Root cause: the payment confirmation arrived after the order had timed out.
  • Fix: re-queue the order. A retry on late confirmations would prevent it.
  • To verify: a ready-to-run query is attached.
📄 check-order.sql

4 min · $0.13 · reviewed before posting

The exchange is illustrative: real tickets carry customer details and stay private.

And it is cheap to run

The expensive model only does what it must. Everything else goes to a far cheaper one.

$0.13per answered ticket
~$16a month at today's volume
~50%of tickets are ops requests it leaves alone
300×cheaper tokens for the investigation

Costs are my own measurements from running the bot.

The rest of the toolkit

Smaller pieces, each built to remove one daily friction.

Shared skills

One set of skills for four AI tools

Coding practices, deploy steps, read-only database access and a morning report. Written once, installed in Claude Code, Codex CLI, pi and opencode.

Live: 12 skills
SDLC

A seven-step path for every feature

Sync, branch, plan and get approval, build, test against the API and the database, write it up, push. It is a skill, so AI tools follow the same path people do.

In progress: team rollout
API collections

1,288 API requests, moved into git

72 Postman collections became 5 Bruno collections beside the code, with a runbook plain enough for an intern or an AI agent to add a service.

Live
Session history

Every AI session, summarised and searchable

Each coding session ends as a short, redacted summary I can search and resume. Raw transcripts never leave the laptop.

Live
Secretariat

A memory for the work itself

People, projects, tasks and reminders in one linked graph on my own Postgres, open to any AI tool over .Model Context Protocol (MCP)An open standard that lets AI assistants call external tools and data sources. It answers what is open, due or waiting.

Live
Story pages

Pitches that read like a story

A design system for one-page pitches, with an honest label on every mock. This site is built with it.

Live
03

After hours, I still build

The habit does not stop at the office door. Two side projects for myself, and before backends, research on encrypting images.

Side project · private repo

gin, one memory for all my AI tools

A personal memory server written in Rust. It speaks MCP, keeps memories in Postgres with ,pgvectorA Postgres extension that stores embeddings, so text can be searched by meaning as well as by keyword. and has a live 3D scanner for looking around inside them.

RustMCPPostgrespgvectorLive
Side project · private repo

A brief waiting for me every morning

A bot that gathers the day's news and my spending, then writes one short brief each morning: what is notable, and how the month's budget is holding up.

Daily briefNewsBudgetLive: runs every morning
Before backends · two papers and a hackathon final

The toolbox so far

Languages
JavaRubyGoPythonRust
Data
PostgreSQLMySQLRedisClickHousepgvector
Messaging
Apache KafkaRabbitMQ
APIs
RESTGraphQLBFF
AI tooling
Claude CodeCodex CLIpiopencodeMCPAgent Skills
Frameworks
Spring BootRuby on RailsDjangoSalesforce Aura and LWC
Education
B.Tech in Electronics and Communication Engineering, SASTRA, 2017 – 2021
04

Three rooms taught me three things

How I got here. Before BondScanner, I learned how software survives rules, growth and real customers, one room at a time.

Lesson one · Rules

First, software that answers to a regulator

DeloitteBusiness Technology AnalystJul 2021 – Feb 2022

My first job was on the State of Colorado's benefits system (CBMS), built on Salesforce. Every feature existed because a regulation said so, and every release had to prove it.

  • Built compliance-driven features in Aura and Lightning Web Components.
  • Raised test coverage across the application's classes to clear Salesforce's release rules, and delivered the project's highest overall coverage.
  • On the side, trained an Einstein Vision image model to estimate the cost of vehicle damage.

What stuckProve it before you ship it.

SalesforceAuraLWCEinstein Vision
release-checklistillustrative
✓
Rule mapped to a featureThe regulation decides what gets built.
✓
Feature builtAura and Lightning Web Components.
✓
Tests cover the classesCoverage has to clear the platform's bar.
✓
Release clearedOnly now does it reach the state's users.
Lesson two · Scale

Then, changing the engine while the car is moving

Good Creator Co, The Good Glamm GroupSoftware EngineerFeb 2022 – Mar 2024

Good Creator Co, the group's creator-marketing platform, ran on a Rails monolith, Java microservices and Go gateways. The business was outgrowing the monolith, so the backend had to move while brands and creators kept running campaigns.

  • Migrated the Rails backend to Spring Boot and Go microservices, with scripts moving huge PostgreSQL and MySQL tables while everything stayed live.
  • Owned the notification and product microservices, wiring in CRMs, Slack, WhatsApp, email and WebEngage.
  • Automated campaigns end to end: creator invites, shortlisting, vouchers, instant payouts, and a planner that forecasts reach and budget.

What stuckScale is a migration you run with the lights on.

RailsSpring BootGoPythonPostgreSQLRabbitMQRedis
illustrative
$ migrate campaigns --to java
→ copying rows in batches
✓ batch 0001 checksums match
✓ batch 0002 checksums match
! replica lag, slowing down
✓ dual-write on old and new agree
✓ reads switched to Java
Lesson three · OwnershipIn progress: re-architecting ULTRA

Then, owning the product and not only the API

Tap InvestSDE 2Apr 2024 – now

At Tap Invest I designed the backend and the Backend-for-frontend (BFF)A thin server layer built for one app. It gathers what a screen needs from many services and returns it in a single response. layer behind two Server-driven appThe server sends the app a description of each screen, so layouts and content can change without a new app-store release. apps, ULTRA by Tap and Tap Bonds. Then I became product owner of the Partner Platform, answering for what it does as well as how.

  • Designed and built the BFF layer that personalises ULTRA by Tap and Tap Bonds for retail investors.
  • As product owner, turned the Partner Platform into an API provider for a range of financial products.
  • Drove backend development for Tap's SaaS platform, and tuned performance, scalability and reliability.

What stuckKnow what the user needs before you draw the boxes.

BFFServer-driven UIPartner APIsSaaS
ULTRA by Tap Tap Bonds Partner apps BFF layershapes each screen PartnerPlatform Backendservices

Tap Invest, simplifiedIllustrative

BondScanner needs all three, on the same day

Each room taught one thing. The platform I architect today needs all of them at once.

Rules

Learned at Deloitte.

  • Regulated by SEBI as an online bond platform
  • Every trade placed on NSE and settled by its clearing corporation
Live

Scale

Learned at Good Creator Co.

  • 1.2 lakh+ investors on the platform
  • ₹125 Cr+ in bonds transacted
Live

Ownership

Learned at Tap Invest.

  • Payments, inventory, trades, settlement, holdings
  • The architecture and the team are mine to answer for
Live
BondScannerneeds all three, at once
05

Starting something from zero? Let's talk.

I am an engineer at heart with a product owner's habits. If you are at the first schema, the first service or the first hires, I would like to hear about it.

LinkedIn GitHub Email me