CV
Arjun Parasuram Prasad
Summary
MS CS @ NYU | SWE Intern @ Corvic AI | ex-Providence, SuperPe
Education
- Master of Science in Computer Science2026-05-01NYU, Courant Institute of Mathematical SciencesGPA: 3.85/4Courses: Machine Learning, Computer Vision, Big Data & ML Systems, GPU Programming, Algorithms, Operating Systems
- Bachelor of Engineering (Computer Science)2023-06-01BITS Pilani — Hyderabad Campus, IndiaGPA: 8.65/10Courses: Deep Learning, NLP, Data Structures & Algorithms, Database Systems, Probability & Statistics
Work Experience
- Software Engineer Intern, AI/ML2026-02-01 -Corvic AIBuilt AI-driven data enrichment pipeline (Web Augment) with Temporal workflows, batched execution, and row-level retry (~2,500 lines across 28 files). Redesigned DataApp embedding API with dynamic dimension resolution via HuggingFace AutoConfig. Implemented full-stack MCP tool integration for AI agent platform (~1,500+ lines across ~20 files).
- Data Analytics Engineer2023-06-01 - 2024-08-01Providence Health CareEngineered ETL pipelines using Azure Data Factory for an in-house data migration tool (Ultrasonic), handling 10M+ row datasets across diverse source/target systems. Partnered with marketing stakeholders to build a Snowflake-based analytics framework; integrated advanced stored procedures with Power BI dashboards to internalize campaign KPI tracking, eliminating >$2M in outsourcing costs. Shipped CI/CD pipelines to automate SQL script deployments, reducing manual effort by ~60 hours/month.
- Machine Learning Intern2023-01-01 - 2023-06-01SuperPe.inBuilt analytical dashboards to track merchant onboarding and financial KPIs; identified bottlenecks that reduced drop-offs by 40%. Developed an OCR pipeline to extract/validate merchant document numbers with 97% accuracy (Google Vision + OpenCV), removing manual entry; deployed via AWS Lambda and Amazon EFS. Created a RAG-based conversational agent (LangChain + Milvus) to automate customer query resolution via agent–tool workflows.
Skills
Languages
- Python, C++, C, SQL, Java, TypeScript, JavaScript
ML / AI Engineering
- PyTorch, Hugging Face, vLLM, DeepSpeed, LangChain, LangGraph, CUDA, OpenCV, scikit-learn, TensorFlow, Search & Ranking, RAG, Vector DBs (Milvus, FAISS), Distributed Training/Inference, LLMs
Data Engineering
- Snowflake, PostgreSQL, Polars, PyArrow, HDFS, MapReduce, Hive, Trino, Kafka, Azure Data Factory, Power BI, ETL, CI/CD
Infrastructure & DevOps
- Docker, Kubernetes (GKE), AWS (Lambda, S3, EFS), GCP (Vertex AI, Artifact Registry), Singularity, SLURM, gRPC, Protobuf, Temporal, FastAPI, Git, BAML, MCP
Publications
- Context-Aware Behavioral Fingerprinting of IoT Devices via Network Traffic Analysis202320th International Conference on Security and Cryptography (SECRYPT 2023)
- Auto-Markup Benchmark: Towards an Industry-standard Benchmark for Evaluating Automatic Document Markup2023Balisage: The Markup Conference 2023
Portfolio
- CoBaLI – Continuous Batching for LLM Inference2025PortfolioDeveloped C++/CUDA inference engine on top of llama.cpp for Qwen2.5-0.5B-Instruct on RTX 4070. Implemented sequential serving, continuous batching, and chunked prefill. Achieved 6.2× throughput improvement (154s→25s) via GPU-assisted scheduling without modifying model kernels.
- EcoScan – Sustainable Fashion Scanner (1st Place HackNYU 2025)2025PortfolioBuilt mobile app scanning clothing for Eco-Scores (0–100) with AI explanations and greener alternatives. React Native/Expo frontend, FastAPI backend, Supabase datastore. Integrated Gemini 2.5-flash, Cloud Vision OCR, and Lykdat API. Won 1st place in HackNYU 2025 Sustainability Track (MLH).
- RepoRecSys – GitHub Repository Recommendation System2025PortfolioTwo-tower neural recommender with 64d embeddings, InfoNCE contrastive loss, cold-start handling via Pub/Sub, and hot model reloading from GCS. Containerized FastAPI inference with <100ms median latency.
- Fine-Tuning Llama 3.2–3B on IPCC Climate Reports2025PortfolioApplied 4-bit quantization, BF16, and LoRA (r=8, α=32) with gradient checkpointing. Orchestrated data/tensor/pipeline parallelism across 2×A100s; doubled throughput (337→169 min) and cut per-GPU memory by ~48%.
- MOT-TR — Fine-Tuning DETR for Moved-Object Detection2025PortfolioFine-tuned DETR for scene-change analysis in image pairs. Reduced eval loss from 2.38 to 0.60 via 3-stage training with progressive unfreezing and cosine annealing.
- Limit Order Book Midpoint Prediction for Crypto2024PortfolioCustom neural-SVR with dynamically scaled ε-insensitive loss and trend-alignment penalty. MSE 28.05, R² 0.998. Benchmarked KNN, SVR, MLP, LSTM, CNN+LSTM, CNN+SVR.
- OS Simulation Suite2024PortfolioFour self-contained OS kernel-concept simulators in C++ covering linking, CPU scheduling (FCFS/SJF/SRTF/RR/Priority), virtual memory (FIFO/Clock/Aging/Working-Set), and disk I/O scheduling (FIFO/SSTF/LOOK/C-LOOK/FLOOK).
- VisaPulse – AI-Powered F-1 Visa Mock Interview2025PortfolioLangGraph state machine orchestrating voice-based mock interviews with RAG over Reddit cases, document analysis, and personalized PDF evaluations via reportlab.