Shivang Bhatnagar

Software Engineer

• Led an AI proof of concept: MCP server for LLM access to MSSQL data plus a 3D avatar; earned CEO's letter.
• Deployed an event-driven S3/SQS/Fargate pipeline on legacy MSSQL data to speed up endorsement validation.
• Led and mentored 2 juniors migrating 10 read APIs to a common API layer on MF Central (2.5M users).
• Moved unit calculations from SQL procedures to a Node.js/TypeScript microservice, cutting peak load 30%.
• Hardened Loan Against Mutual Funds APIs for VAPT security testing: request validation, rate limiting.
• Built Group SIP, a payroll-based investment platform: 30+ APIs, 95% coverage; MongoDB cut latency 40%.

Software Engineer Intern

• Re-engineered backend by replacing 8+ SQL procedures with Node REST APIs, reducing response time by 90%.
• Established quality assurance by writing Jest unit tests, which reduced post-release defects by 55%.
• Delivered real-time Loan Against Mutual Funds dashboard in React.js for clients and cross-team visibility.
• Authored OpenAPI (Swagger) documentation for 13+ API endpoints, streamlining developer onboarding.

Frontend Web Development Intern

• Engineered dynamic, data-driven email templates using Handlebars.js for personalized communications.
• Developed a Node.js automation tool converting CSV data to PDF reports, reducing processing time by 60%.
• Streamlined asset tracking for CRIS (Centre for Railway Information Systems) via automated tagging.
• Integrated modular React components into an Electron-based desktop application, working with other teams.

• Dropped the 36 s AI-image step: text search takes 0.3 s and ranked products 2.6x better (20 queries).
• Added failover (0 errors, one server down) and load shedding: 4 req/s at 32 parallel requests, was 0.
• Engineered idempotent SQS ingestion with retries & DLQ replay; systemd restarts failed services in <8 s.
• Defined 4 AWS CDK stacks (IaC) with CloudWatch request tracing and 10 alarms; clean teardown, 0 left.
• Shipped on Cloud Run (Firestore) with rate limits, SSRF allow-list and a safety guard; 360 CI tests.

• Python
• TypeScript
• AWS (EC2, SQS, CDK)
• Google Cloud Run
• SigLIP
• Qwen3
• Ollama

• Built a 4-layer LLM cache (exact, semantic, retrieval, KV) for RAG; semantic layer cut Claude API cost 39%.
• Evaluated recall@k, F1, p50-p99 latency and wrong-answer rate on Qwen and Claude via a harness, 40+ runs.
• Eliminated stale answers after data edits via data-version invalidation (29 of 30 stale before, 0 after).
• Counted local KV-cache reuse: 25–34% of prompt tokens vs 2% cold; Claude prompt caching cost 18–25% more.
• Deployed a 1.62 GB arm64 Docker image to AWS ECS Fargate with API-key auth, spend cap and health check.

• TypeScript
• Node.js
• SQLite
• Docker
• AWS (ECS Fargate)
• Claude API
• Ollama
• Vitest

B.Tech, Computer Science Engineering · GPA 3.4/4