I build software from stakeholder requirements through delivery. Previously at Jio Platforms and founding engineer at GoldJar; now pursuing my MS in Software Engineering at Carnegie Mellon University.
Aug 2026 – Expected May 2027 · Silicon Valley
2019–2023 · Grade: 9.31
Select a focus to explore my toolkit.
Application services and business workflows.
Retrieval, tool use, and evaluation.
From understanding a business to shipping software.
Professional experience. Clear ownership.
Results grounded in the work.
Bringing offline bullion operations online.
Visit product ↗I worked directly with business owners to translate gold-wholesale operations into application logic, building GoldJar end to end alongside my full-time Jio role.
The platform now supports over US$1 million in daily gold and commodities trading volume. The business has reached profitability and onboards multiple clients each day.
Trading volume represents transaction value, not revenue.Walk through a real booking from user action to order state. Explain which business rules you implemented, how pricing was represented, and how you tested changes.
These are discussion prompts, not claims that a specific architecture was used.
An internal application for Reliance Retail.
In the Jio Matrix team, I worked with stakeholders to translate incentive-management requirements into an internal application and owned delivery end to end.
Requirements translation, application implementation, testing, and delivery of the incentive-management workflow.
Reduced manual effort and improved productivity by 90% for the workflow.
The figure describes workflow productivity, not a company-wide improvement.Prepare a sanitized example of an incentive rule, how you clarified it with stakeholders, and how you validated the application against the original process.
Built real-time features and AI assistants using Python, LangGraph, and LangChain. Worked on tool calling, retrieval, response evaluation, and system performance.
Helped scale online meetings from 50 to 5,000 participants and reduce infrastructure costs by 25%.
Independent implementations grounded in
AI quality, backend reliability and fintech.
A response evaluation workbench for citation validity, expected facts, abstention and latency budgets.
Explore at-least-once delivery, idempotency conflicts, exponential backoff and dead-letter handling.
Match split settlements against orders and surface missing, duplicate, orphan and mismatched transactions.
New portfolio projects built with AI assistance. Each runs locally in your browser with synthetic data. Explore the implementation and tradeoffs inside each project.
New portfolio demonstrations.
Synthetic data. Runs in your browser.
Submit a synthetic bullion booking, retry the same request, and inspect the ledger. Integer cents avoid floating-point money calculations. An idempotency key prevents a retry from booking twice.
Ready. Submit the same request twice to test a retry.
A synthetic incentive policy with explicit thresholds. Adjust sales and returns, then inspect each calculation step. This is a new demonstration, not Reliance’s policy or internal application.
Policy: below 80% target → 0%; 80–99.99% → 1%; 100–119.99% → 2%; 120%+ → 3%. Rate applies to net sales. Payout capped at $2,500.
Search a small synthetic operations handbook. The workbench ranks documents by token overlap, cites matching passages, and abstains when it finds no evidence.
Deterministic lexical retrieval, not an LLM or production RAG service. No query leaves your browser.
Three browser-local modules: idempotent booking with balanced journal entries; a deterministic rules engine with a calculation trace; and a document retriever with citations and abstention. Booking data lives in memory and resets on reload. No external APIs or credentials.
A database transaction and unique idempotency constraint; authorization; durable audit logs; concurrency tests; versioned policies; and retrieval evaluations. These are next steps, not deployed backend capabilities.
These demonstrations are independent educational implementations. They do not reproduce employer source code, confidential rules, or production systems.
Karthik Sridhar & Pranita Mahajan · CCIS 2440 · 22 May 2025
Compared a random baseline, Logistic Regression, Linear SVM, and XGBoost for duplicate-question identification using Matthews correlation coefficient.
Samundiswary S, Karthik Sridhar, Mehul Aswar & Simran Kaur Shembe · IEEE Xplore, 16 October 2023
A LeNet convolutional neural network for German street-sign classification, paired with text-to-speech alerts.
I enjoy working where engineering meets business operations: understanding a workflow, finding the rules that matter, and turning them into usable software.
I’m exploring 2027 new-grad opportunities and internships in backend software engineering and AI-agent engineering.
Get in touch ↗MS Software Engineering · Silicon Valley
Expected May 2027
BE Information Technology · 2019–2023
Grade: 9.31 · Teaching Assistant, Database class