Charting a course from raw data to working systems — deep learning models, AI agents, and pipelines that hold up in open water. M.S. Data Science, Stevens Institute of Technology. AWS certified, twice over.
I'm Aditya — a data scientist who likes the whole voyage, not just the destination: framing the question, wrangling the data, training the model, and shipping the thing.
I trained as a computer engineer in Mumbai, then crossed an ocean for an M.S. in Data Science at Stevens Institute of Technology — overlooking the very harbor this page is drawn from. Along the way: speech-to-speech translation systems, diabetic retinopathy classifiers over 143k images, and AI agents on AWS Bedrock.
Daily driver: Arch Linux. Current obsession: multi-agent RAG systems. Long-term heading: AI systems that are useful in production, not just in notebooks.
Every port taught the crew something different — tuning, pipelines, and how to keep models honest in production waters.
Each island is a shipped project. Drop anchor on any of them for the full chart.
A real-time supplier quality and capacity monitoring platform, built the way it would ship in production: streaming ingestion, orchestrated pipelines, edge computer vision, and an agent that explains what broke.
End-to-end platform for monitoring global supplier production performance across quality, capacity, and equipment health. Streams telemetry through FastAPI and gRPC into Kafka, orchestrates transformation with Airflow, and enforces data contracts with Great Expectations before anything lands in TimescaleDB. A YOLOv8 model runs defect detection on the edge, and a LangChain agent performs root-cause analysis over the resulting time-series so operators get an explanation rather than just an alert. Fully containerized with Docker Compose and wired to GitHub Actions CI.
A "jury" of four models that catches AI-generated insurance fraud. Vision and reasoning are deliberately kept apart so neither can contaminate the other's judgment.
Built in 48 hours and took first place. The problem: telling genuine insurance claim evidence apart from AI-generated fakes across images, PDFs, and video. The approach I hadn't seen tried was separating the two halves of the decision — a Vision Agent (Qwen-VL-Plus) performs pure forensic analysis on keyframes extracted with OpenCV, looking only at physical and structural anomalies. Its findings then pass to three independent Critic Agents (Qwen Turbo, DeepSeek R1, GLM 4.6) that reason over the evidence without ever seeing the image, and vote. Fraud is only flagged on consensus, which removes single-model bias. Includes a Dashcam Override Protocol that distinguishes genuine motion blur from the unnaturally smooth surfaces of generated footage, plus SQLite-backed batch processing for claim backlogs. Validated against a custom 500-file dataset of real and AI-generated claims.
A low-latency, concurrent translation pipeline on a producer–consumer model: Whisper for STT, MarianMT for translation, Meta MMS-TTS for the voice on the other side.
Architected a low-latency, concurrent translation pipeline using producer-consumer model with Python threading, achieving ~2.8s pipeline depth. Integrated Whisper Medium for STT (4% WER) and MarianMT for translation, achieving BLEU scores of 80.85 (EN→ES) and 71.64 (EN→FR). Implemented custom VAD using Silero with <3ms frame latency and engineered hallucination filters.
Is free market data good enough? A reproducible benchmark scoring Yahoo Finance against CRSP and Compustat on the dimensions that actually cost you money.
Built a one-command analytics pipeline comparing Yahoo Finance against the institutional gold standards CRSP and Compustat (via WRDS). Implemented formal data quality scoring with configurable thresholds, then quantified the downstream impact through risk-adjusted return metrics, factor exposure diagnostics, and scenario sensitivity analysis. Outputs a recommendation matrix mapping each provider to the use cases it can be trusted for, surfaced through a Streamlit dashboard and covered by pytest smoke tests.
A Chrome extension that reads your resume and fills job applications across Workday, Greenhouse, LinkedIn, Indeed, and iCIMS — without tripping their bot detection.
Manifest V3 Chrome extension that parses resumes in PDF, DOCX, DOC, and TXT, then auto-fills job application forms across five major applicant tracking systems. Combines deterministic regex extraction with LLM parsing through OpenRouter, OpenAI, and Gemini, keeping a human review step before anything is submitted. Solving the hard part meant React-safe field filling that triggers the framework's own event handlers, plus anti-bot-aware input emulation so keystrokes register as genuine.
Deep learning severity grading over 143k retinal images, with PySpark handling the heavy lifting on Azure Blob Storage.
Processed a large-scale image dataset (143,509 images, ~22GB) using Apache Spark with Azure Blob Storage for distributed access. Engineered a deep learning pipeline using custom VGG-16-based CNN, achieving 87%+ accuracy in identifying diabetic retinopathy severity levels. Reduced training time by 40% using Spark distributed compute.
A LangChain agent on AWS Bedrock that gives grounded career guidance instead of confident nonsense.
Architected an autonomous chatbot using LangChain, AWS Bedrock (Claude V2), and Streamlit with Agentic AI concepts like goal-seeking and tool-use. Developed a RAG pipeline that increased response relevance by 40% by integrating external career data. Designed an alignment & safety framework that reduced model hallucinations by 75%.
CrediSync — a dual-agent system that reads hospital credentialing PDFs and fills the forms for you.
Developed CrediSync, a multi-agent automation system for hospital credentialing, reducing manual form-filling time by ~90%. Engineered dual-agent architecture with LangChain: Parser Agent extracted 15+ key entities from PDFs, Form Filler Agent achieved 100% field mapping accuracy. Built custom Chrome extension for one-click auto-filling.
A deep learning sentiment pipeline over 1.6M tweets, benchmarked against classical models to prove the lift.
Preprocessed 1.6 million tweets using tokenization, stopword removal, and stemming. Implemented deep learning pipeline using Sequential Model with Embedding, Conv1D, MaxPooling1D, and LSTM layers, achieving 85%+ accuracy. Compared SVM, Decision Tree, Bi-LSTM, and BART models, achieving 12-18% higher F1-score with deep learning approaches.
Research into neural data compression for images and video — squeezing files without losing fidelity.
Working on Neural Data Compression for Images and Videos, exploring state-of-the-art deep learning techniques for efficient media compression while maintaining quality. Building foundation in AWS cloud services for scalable deployment.
Everything in the hold has been used on a real crossing — no decorative cargo.
GPA 3.7 · NLP, Statistical Methods, Time Series Analysis, Database Management, Deep Learning. Thesis-adjacent tinkering: multi-agent RAG systems and neural data compression for high-fidelity images and video.
CGPA 8.37 · Scalable software architecture, algorithm design, and core programming — plus the hackathons that pulled me toward machine learning in the first place.
Took first place with a Vision-Critic pipeline for insurance fraud detection — four models acting as an independent jury, with visual forensics deliberately isolated from logical deduction so neither could contaminate the other. Built in 48 hours and held up under the judges' stress-testing.
Hiring, collaborating, or just comparing charts — the radio is always on. I usually reply within a day.
