Ayush Varma
Contact
Seeking full-time software engineering roles Get in touch

Portfolio

Systems that hold up
under load.

I am a software engineer working on distributed systems, LLM evaluation infrastructure, and performance engineering. Most recently a Software Developer Intern at IBM, where I built the benchmarking platform that stress-tests AI coding assistants on mainframe modernization. Before that, two years at Oracle on State Street's exchange platform, a system clearing more than ten million transactions and millions of dollars in US fixed income instruments every day.

Master's student in Computer Science at Northeastern University, specializing in programming design paradigms and machine learning operations. I care about making large, complex systems feel simple from both sides: interfaces the people using them never have to think about, and an architecture the next engineer can scale and extend without a walkthrough. And I am drawn to the part of engineering where a number gets measurably better and you can prove it.

Selected work Download resume

At a glance

Now
MS Computer Science, Northeastern University
Most recently
Software Developer Intern, IBM
Before that
Software Developer, Oracle Financial Services
Based
Boston, Massachusetts
Focus
Distributed systems, LLM evaluation, performance engineering

Selected figures

Four results, with the work behind them

Oracle / Exchange Platform

10M+

Transactions cleared daily

Built the modern data platform behind State Street's exchange platform, migrating it off on-premise Hadoop onto cloud-native AWS Databricks. It clears millions of dollars in US fixed income instruments every day.

Oracle / Data Lake

790+

Tables migrated

Part of a 22TB data lake transformation. Partition strategy, Z-ordering and caching were tuned per table so query latency held up across the whole migrated estate rather than degrading with scale.

IBM / Evaluation Service

512

Concurrent users held

The service used to fall over at roughly four concurrent users. After bounding resident memory it completed the full ramp to 512 users and 30,660 requests with a 0.00% error rate.

IBM / Hottest Endpoint

61×

Less peak memory

Swapping a full-table scan for cursor pagination and server-side streaming took the heaviest endpoint from 13.4 GB resident at 100% errors down to a flat 220 MB at zero.

01 / Profile

Backend depth, with a bias toward evidence.

What I work on

Distributed systems, data platforms, and the evaluation infrastructure around large language models. Most of my work has been in enterprise contexts where the failure modes are load-dependent and the fix has to be provable.

How I work

I build the measurement before I build the fix. The memory work at IBM started as a harness that could reproduce a crash on demand, and that harness is what turned an intermittent outage into a one-line root cause.

Design discipline

I studied programming design paradigms formally and use them deliberately: SOLID boundaries, the Command pattern for undo, Strategy for pluggable exporters, and test suites strong enough to survive mutation testing.

Currently

Finishing a Master's at Northeastern in Boston, after a summer at IBM in San Jose. Open to full-time software engineering roles.

02 / Experience

IBM

May 2026 → Aug 2026

Software Developer Intern / San Jose, California

Worked inside an internal LLM evaluation platform that stress-tests IBM's AI coding assistants on mainframe modernization tasks: code explanation, COBOL to Java transformation, and generation.

  • 01

    Extended the benchmarking pipeline builder, pairing a Python and Streamlit wizard flow with an evaluation service over HTTP.

  • 02

    Operationalized an LLM-as-a-Judge layer that grades generated code through rubric scoring, pairwise A/B comparison and risk assessment, replacing subjective manual review with reproducible grading at thousands of samples per run.

  • 03

    Unified two assistant surfaces behind one benchmarking flow, refactoring roughly 3,000+ lines of near-duplicated logic into a single shared module.

  • 04

    Root-caused an out-of-memory crash under load with a reproducible harness that ramps concurrency while sampling process memory in real time, isolating it to unbounded query deserialization, then re-architected the hottest endpoints to bound resident memory.

5.5×

Less peak memory

6.0 → 1.1

Gigabytes resident

100 → 0

Percent error rate

4 → 512

Concurrent users

READ THE FULL TECHNICAL BREAKDOWN

Northeastern University

Jan 2026 → Present

Assistant Data Analyst / University Decision Support

  • 01

    Engineering automated Destination Survey pipelines in Python with Selenium to track post-graduation outcomes for the Provost's Office.

  • 02

    Built a Confidence Scoring System that weights reliability across self-reported surveys and automated scrapes.

  • 03

    Architecting Tableau Prep ETL workflows for strategic program evaluation and accreditation, with data quality validation scripts supporting enrollment and student success analytics.

Oracle

Jul 2023 → Aug 2025

Software Developer / Financial Services, Bangalore

  • 01

    Designed scalable microservices in Java and Spring Boot for an AI-powered meeting insights platform, presented independently to an Oracle Group VP and deployed to production.

  • 02

    Migrated a financial platform handling 10M+ daily transactions from on-premise Hadoop to cloud-native AWS Databricks with fault-tolerant architecture.

  • 03

    Optimized query processing across 790+ migrated tables with partition strategies, Z-ordering and caching in a large-scale data lake.

  • 04

    Automated pipelines with Apache Airflow, moving from schedule-driven to event-driven execution with file and task sensors, cutting manual intervention by 35%.

READ THE FULL TECHNICAL BREAKDOWN

Eastern Software Systems

Jan 2023 → May 2023

Software Developer Intern

Designed the Parliament of Kenya's Document Management System with role-based access control, version control and distributed caching, cutting document retrieval time by 35% while supporting 500+ concurrent users daily. Built 15+ reusable microservice components following object-oriented design principles.

03 / Selected Work

ALL REPOSITORIES →

01

AskBuddy

Python · RAG · pgvector · Slack

+

A Slack bot that routes workplace questions across five specialist agents, answers from your own HR and IT documents with the exact filename and effective date cited, and keeps GitHub issue and PR activity visible. When the documents do not contain an answer it says so instead of guessing.

Notes

  • Supervisor routes to five agents: HR, IT, scheduler, GitHub issues, GitHub PRs.
  • Separate corpora, so an HR question never returns IT content.
  • Recurring reminders persist in Postgres and survive restarts.
REPOSITORY DOCS

02

Mellow

Android · Java · Firebase

+

An Android wellness app built on the idea that consistency beats intensity. Sub-minute guided breathing, mood tracking with on-device insights, journaling, and home-screen widgets, reinforced by streaks with a forgiving grace period and a coin jar that fills over time.

Notes

  • Full design system: palette, typography, iconography, mascot, and every screen, documented as an interactive page.
  • Five-colour mood scale resolves consistently across calendar, journal, charts and widgets.
  • Mascot built as a VectorDrawable, no external art or Lottie.
REPOSITORY DESIGN SYSTEM

03

Calendar Engine

Java · Swing · JUnit · PIT

+

A multi-calendar management system built as an exercise in design discipline. Recurring events with complex weekday patterns, three-tier edit propagation, conflict detection, and IANA timezone conversion for global scheduling.

Notes

  • Interface Segregation keeps GUI and CLI clients talking to the engine through contracts only.
  • Command for undo and redo, Builder for construction, Strategy for pluggable CSV and iCal export.
  • 90%+ mutation test coverage under PIT, so the assertions are verified, not just the lines.
REPOSITORY DESIGN DOC

04

SportsTV Warehouse

MySQL · ETL · Kimball

+

A Kimball star schema for real-time streaming engagement analytics, with ETL pipelines that processed 1.18 million transactions at roughly 23,000 per second.

Figures

Transactions 1.18M
Throughput 23K/sec
REPOSITORY LIVE DOCS

05

Restaurant DB

MySQL · R · OLTP

+

A cloud-hosted normalized database for restaurant chain analytics, migrated to third normal form from 179,000 denormalized point-of-sale records, with ETL pipelines, stored procedures and business intelligence reporting on top.

Figures

Source records 179K
Target form 3NF
REPOSITORY ANALYSIS

06

InvestMate

Spring Boot · Oracle SQL · REST

+

A full-stack mutual fund investment platform with investor and manager dashboards, portfolio tracking, buy and sell execution with atomic transactions, dynamic Net Asset Value calculation and wallet management.

Recognition

National 1st runner-up at Oracle's GradHire Hackathon.

Fund composition management and investor participation tracking cut manual tracking errors by 30%.

REPOSITORY DESIGN DOC

04 / Stack

Languages

Java · Python · C++ · TypeScript · JavaScript · SQL

Frameworks & Tools

Spring Boot · FastAPI · Streamlit · React · Node.js · REST APIs · Git · Jenkins · CI/CD

Data & Cloud

AWS (EC2, S3, Lambda, Databricks) · Oracle Cloud Infrastructure · Apache Spark · Apache Airflow · MySQL · PostgreSQL · pgvector

Systems & Architecture

Microservices · Distributed Systems · Object-Oriented Design · Design Patterns · Performance Engineering · LLM Evaluation

05 / Education

Northeastern University

Aug 2025 → May 2027

Master of Science, Computer Science / Boston, MA

Coursework: Programming Design Paradigms, Database Management Systems.

Shiv Nadar University

Jul 2019 → Jun 2023

Bachelor of Science, Computer Science

Coursework: Data Structures and Algorithms, Software Engineering, Object-Oriented Programming, Applied Linear Algebra, Cloud Computing.

Certifications

AWS Certified Cloud Practitioner
Oracle Certified Foundations Associate

Leadership

Webmaster and Organizing Committee, ACM Shiv Nadar. Ran Hackdata 5.0, a 48-hour multi-track hackathon with 400+ participants across 80+ teams.

06 / Contact

Let's talk about hard problems.

Open to full-time software engineering roles. Based in Boston.

Email ayushvarma04@gmail.com LinkedIn /in/ayushvarma7 GitHub @ayushvarma7 Resume Download PDF