Building a multi-agent workspace to streamline client proposal development — covering institutional knowledge retrieval, expert search, deep research, and competitive intelligence. Shipping it company-wide as a Claude-native plugin and MCP server integration, cutting proposal turnaround by 20+ hours and freeing teams to focus on strategy and pitch quality over research grunt work.
Hey, I'm Swastik
swastik3[at]umd[dot]edu
Hi, I'm Swastik Agrawal. I'm an undergrad at the University of Maryland, College Park studying CS and Math. I spend most of my time doing machine learning research — specifically around mechanistic interpretability, AI safety, and deep learning for scientific applications.
I am currently working on Applied AI at BCG X (6th internship). I've Interned at 4 other companies in the past and have been working on AI products, frameworks and tools for 2 years now.
I also co-founded AI/ML at UMD, a 600+ member student organization. Quite proud of the community we have been able to build here.
Outside of work, I am an avid hacker, and used to participate in multiple hackathons every month. I've been to ~15 hackathons (7 wins: MIT, Harvard Hackerhouse, Georgia Tech, UMD, etc). Love playing just about any sport (especially squash/volleyball/cricket).
Research
My current research spans mechanistic interpretability and AI-generated text detection. In Dr. Sarah Wiegreffe's group at UMD (work accepted at the ICML 2026 Mech Interp Workshop), I'm investigating the performance gap in activation steering — specifically Contrastive Activation Addition (CAA) — controlling for steering coefficient, layer, token position, and dataset formulation. I'm also exploring a geometric proxy using the difference-of-means line to estimate test-time steering success across multiple-choice and open-ended generation settings.
Through the Supervised Program for Alignment Research (SPAR) (also accepted at the ICML 2026 Mech Interp Workshop), I developed Circuit Oracle — an agentic pipeline for automated end-to-end circuit discovery, interpretation, and causal hypothesis generation in LLMs. This includes conducting jailbreak circuit discovery on Qwen-3-4B and Gemma-2-4b-it against GCG and AutoDAN attacks.
In Dr. Fardina Alam's group (work accepted at COLM 2026), I co-led research on training-free AI-generated text detection using signals from letter distributions. We built the MDTA benchmark — 642,274 prompt-aligned samples across 4 models, 5 domains, 3 temperatures, and 3 adversarial strategies including lipogrammatic attacks — significantly expanding HC3. We also introduced LD-Score, an orthogonal detection signal that improves AUROC and F1 when integrated with perplexity-based detectors like Binoculars and DNA-DetectLLM.
Previously, I spent over a year in the Risk-Informed Solutions in Engineering Laboratory where I engineered a temporal deep-learning model to impute unavailable data on tropical cyclones, improving existing shallow approaches by 15%. That work was presented at ICOSSAR 2025, the AGU Fall Meeting 2024, and the UMD CEE Symposium 2024. Read the paper.
I've also worked in the Tubaldi Lab, designing a bimodal object classification model integrating temporal pressure sensing and visual input, and developing a communication framework for the UR3e robotic arm with automated trial randomization, sorting operations, and streamlined data collection scripts.
Work Experience
Built a competitor intelligence system with real-time performance alerts to inform social media marketing decisions. Developed a fully end-to-end agentic video creation microservice using Mastra, handling the entire pipeline from content planning through final delivery across 15+ video formats. Managed cloud infrastructure for these features with Terraform across AWS Step Functions, Lambda, ECS Fargate, and S3.
Built an in-house graph framework for LLM pipelines to power multi-threaded graphical workflows with tool calling and SSE response streaming, cutting user latency by ~90%. Migrated the codebase from multiple distributed services to a monorepo architecture, reducing operational costs by ~40%, and enhanced database schemas to enable feature improvements across 4 company products. Refined signup and payment funnels through A/B testing, driving ~80% increase in registration rate. Deployed AWS Lambda functions for audio transcription and file processing, and contributed 500+ unit tests across the codebase.
Developed 10+ agents with sophisticated LangGraph workflows equipped with web search capabilities, compiling travel options through natural language-derived API calls to multiple third-party services.
Migrated the Node.js company portal from Node v12 to v16 incrementally, refactoring code and reducing initial load time by 12%. Implemented an analytics dashboard to monitor tenant usage across enterprise deployments for real-time metric analysis.
Community
Co-Founder & Head of Tech
Jul 2024 – Aug 2025AI/ML at UMD · University of Maryland
Grew the organization to 600+ members and organized 5 workshops with 5+ company collaborations. Served as a UMD hackathon judge and conducted workshops on RAG, LLMs, and finetuning. Managed the development of a recommendation engine for WISE Cities LLC using composite scoring, matrix factorization, and keyword search. Led a research project building a real-time RAG pipeline with lecture transcriptions to answer queries using a dynamically updated knowledge base (piloted in 4 UMD classes).