Ishan Misra
Thinking in Systems. Creating with code.
Overview
Ex Intern @Microsoft, xAI
Social Links
About
I'm a final-year B.Tech student in Computer Science and Engineering (AI) at Manipal Institute of Technology, Bengaluru. I'm deeply interested in software development, especially backend systems and infrastructure; the kind of work that focuses on how things actually run, scale, and stay reliable behind the scenes.
Most recently I interned at Microsoft on Security Copilot, where I owned long-term memory for AI agents end to end, covering extraction, storage, retrieval, and benchmarking. That work lifted extraction F1 from 42% to 91% and cut retrieval latency by up to 65%. Before that I worked with xAI, early-stage startups, and NGOs, building real features, automating workflows, and shipping production code. These experiences taught me how software is built in the real world, with constraints, deadlines, and users who actually depend on it.
Outside of academics, I'm a regular at the gym and a state-level football player. I've also won multiple gaming events during school, which sharpened my competitive mindset and problem-solving skills. Leadership has always been a part of my journey. I served as the Internal Affairs Director in my school's tech and development team, where I helped organize technical initiatives and fostered my interest in building things from a young age.
I enjoy working on systems that are simple on the surface but complex underneath, and I'm constantly learning to become a better engineer. Right now, I'm focused on strengthening my fundamentals, building meaningful projects, and preparing for strong SDE roles.
GitHub Contributions
Education
- Data Structures
- Algorithms
- OOPS
- DBMS
- Machine Learning
- Artificial Intelligence
- Computer Vision
- Natural Language Processing
- Java
- C++
Stack
Experience
- Owned long-term memory for Security Copilot agents end to end across extraction, storage, retrieval, and benchmarking, shipping a production 6-stage extraction pipeline with salience filtering, classification, confidence scoring, and bucketized storage
- Lifted extraction F1 from 42% to 91% over the team's prior one-shot approach, clearing 19/20 adversarial and 43/50 functional test cases against 2/20 and 17/50 for the incumbent
- Benchmarked on LoCoMo, LongMemEval, and BEAM plus 50+ purpose-built synthetic evaluation conversations, scored across Precision, Recall, F1, nDCG, and MRR
- Architected storage on Azure Cognitive Search indexing over CosmosDB, and a hybrid retrieval path pairing ACS vector search with LLM reranking; shipped it behind a flag-gated API in Memory Service
- Cut retrieval latency up to 65% (20–30s to 7–15s) via parallel processing and resolver optimization; load-tested and authored the scale-out HLD (background workers, push vs. pull) the team carried forward into Project Perception
- Azure Cognitive Search
- CosmosDB
- Vector Search
- LLM Reranking
- Benchmarking
- Distributed Systems
- Python
- LLM Evaluation
- Red Teaming
- Adversarial Testing
- Benchmarking
- Automation
Freedom From Poverty Foundation
- Next.js
- React
- Razorpay
- OAuth
- Role-based Access Control
Freelance
- Automation
- API Integration
- Scripting
- Monitoring
- Error Handling
Maidaan
- React
- Node.js
- Gemini API
- Firebase
- Real-time Systems
- Debugging
TechNovel
- SEO
- Content Strategy
- Data Analysis
- Technical SEO
Projects(4)
Multi-agent pipeline that turns raw Jira stories into runnable test suites
- Converts raw Jira stories into runnable test suites, with automated edge-case generation covering branches and failure paths rather than only the happy path
- Aggregates outcomes into a deployment confidence score, wired into CI/CD so execution and reporting run on every build
- Python
- JavaScript
- PowerShell
- CI/CD
- Multi-agent
The unified developer inbox for news, alerts, and system design
- Aggregates tailored developer news, dependency vulnerability and end-of-life (EOL) alerts directly from connected repositories
- Delivers daily system design questions and company radar tracking to help engineers stay interview-ready and informed
- Built with a responsive Next.js frontend and robust API integrations to process and categorize real-time data efficiently
- Next.js
- Supabase
- Tailwind CSS
- shadcn/ui
- TypeScript
- Vercel
- GitHub API
Modular SaaS learning platform with authentication and billing
- Architected a modular system separating authentication, billing, and core learning services
- Implemented secure authentication and subscription-based access using third-party providers
- Built low-latency real-time interaction workflows with ~300ms response times
- Added schema validation and centralized error tracking, reducing debugging time by 50%
- Next.js
- Supabase
- Clerk
- Stripe
- shadcn/ui
- Vapi
- Zod
- Sentry
Personal health command center for tracking wellness metrics and trends
- Built a responsive dashboard to log health metrics, surface trends, and manage daily routines
- Designed a guided onboarding flow to capture user profile details, tracked metrics, and goals quickly
- Implemented wearable data sandboxes (Google Fit, Fitbit, Apple Health) with dummy datasets and CSV export
- Developed rule-based analytics to highlight anomalies, trends, and weekly focus areas
- Added goal tracking for weight, steps, sleep, hydration, and exercise with clear visual indicators
- Integrated charts for multi-series trend visualization and weekly email reports via server actions
- Next.js 15
- TypeScript
- Tailwind CSS
- shadcn/ui
- Framer Motion
- Firebase
- Clerk
- Recharts
- Nodemailer