Research themes
What I focus onResponsible AI and safety, interpretable reasoning, structured knowledge extraction, and evaluation methods that catch “silent failures.”
MS CSE @ UCSC
Safety
NeuroSymbolic
Knowledge Graphs
Evaluation
Transparent, interpretable, human-aligned AI.
I build systems that ▍
My research goal is to combine neural learning with explicit structure and principled evaluation. Across foundation model safety, neuro-symbolic reasoning, structured knowledge extraction, and scientific ML, I’ve seen strong models fail in subtle, opaque ways under deployment and stress. I focus on auditing behavior, understanding representations, and enforcing semantic + reasoning consistency instead of relying only on output-level metrics.
About
I like systems that are simple on the surface and rigorous underneath: clear baselines, measurable progress, and explanations that survive real usage.
Tooling
How I buildPractical LLM/KG systems, evaluation harnesses, and reproducible research code. Comfortable across Python/Java/SQL and modern ML stacks.
Skills
Research engineering + production ML systems.
Languages
Java, Python, C/C++, SQL, JavaScript, HTML/CSS, R
Libraries & Frameworks
PyTorch, TensorFlow, Transformers, LangChain, ChromaDB, XFormers, NumPy, Pandas, Scikit-learn, Statsmodels, Keras, Matplotlib, Seaborn, PyPDF
Cloud & Infrastructure
AWS (EC2, S3, IAM, Lambda, SageMaker, CloudWatch), Docker, Kubernetes, Linux, Git, CI/CD, REST APIs, Nginx
MLOps & Deployment
FastAPI, Flask, ML pipelines, experiment tracking, model versioning, data validation, monitoring & logging, distributed training, GPU optimization, batch & real-time inference
Data Engineering
ETL pipelines, SQL optimization, schema design, data cleaning & transformation, time-series preprocessing, vector databases (ChromaDB), caching strategies
CV Timeline
Student Researcher - AIEA Lab @ UCSC
Sep 2025 – Dec 2025
Independent study: NeuroSymbolic system for analyzing cognitive decline risk from long-term AI usage.
- Integrated neural pattern recognition with symbolic reasoning for transparent insights.
- Building a configurable framework with fuzzy + temporal logic and a live web demo for rule visualization.
AI Research Intern - AI Institute of South Carolina (remote)
Jun 2024 – May 2025
Multimodal safety + bias mitigation across generative and language models.
- Researched multimodal toxicity in VLMs and proposed safer generative approaches.
- DE-HATE: diffusion-based detoxification for hateful AI-generated content.
- Applied machine unlearning (DPO) to mitigate gender, ethnic, and LGBTQ+ bias in LMs.
- Benchmarked time series foundation models for anomaly detection in constrained environments.
Data Science Research Trainee - Space Applications Centre (ISRO)
Jan 2024 – May 2024
Satellite navigation forecasting pipeline focused on reliability and real-world integration.
- Built a custom LSTM pipeline for satellite clock bias prediction, outperforming traditional methods.
- Created preprocessing (10-min resampling, outlier filtering) for time-series stability.
- Integrated predictions into IoT algorithms to improve accuracy and power efficiency.
Student Researcher (AI) - Indus Institute of Technology
May 2022 – Jan 2024
Knowledge graphs from LLMs and a retrieval-augmented medical assistant.
- KG generation pipelines using GPT-4/GPT-3, LLaMA 2, and BERT for semantic extraction.
- Built Med-Bot with LangChain + ChromaDB + LLaMA + AutoGPTQ for grounded responses from medical PDFs.
Papers
Selected publications and preprints.
Leveraging LSTM for Predictive Modeling of Satellite Clock Bias
IEEE XploreLSTM-based clock-bias forecasting with strong accuracy gains and a practical preprocessing workflow for navigation reliability.
Generating Knowledge Graphs from Large Language Models: A Comparative Study of GPT-4, LLaMA 2, and BERT
arXivKG generation directly from unstructured text, evaluated with structural + semantic metrics for GraphRAG readiness.
Med-Bot: An AI-Powered Assistant to Provide Accurate and Reliable Medical Information
arXivRetrieval-augmented medical chatbot using modern LLM tooling to answer questions grounded in medical literature PDFs.
Time Series Foundational Models: Their Role in Anomaly Detection and Prediction
arXivCritical evaluation of TSFMs for anomaly detection/prediction: when they help, when classic methods win, and why.
Projects
Selected builds from my recent work.
Research Forge
Research agentSelf-improving research agent that discovers papers, builds structured research memory, generates hypotheses, and adapts strategy across iterative research cycles.
AI Meeting Assistant
Desktop copilotElectron desktop meeting copilot that captures system audio, transcribes speech, and generates context-aware AI responses with optional screenshots and file grounding.
NeSy AI Cognitive Analyzer
Live demoNeuroSymbolic framework for cognitive-risk analysis with rules, temporal logic, and an interactive demo.
TSFM Evaluation
BenchmarkingBenchmarking time series foundation models for anomaly detection under realistic constraints.
Knowledge Graphs via LLMs
ToolkitKG extraction + evaluation pipeline for structured reasoning and GraphRAG readiness checks.
Satellite Clock Prediction
ISRO
Clock-bias forecasting pipeline with robust preprocessing and deployment-minded evaluation.
Note: This project cannot be open sourced due to security restrictions by the Indian Space Research Organization.
Medical AI Chatbot (Med-Bot)
RAGRAG chatbot grounded in medical PDFs with efficient inference tooling.
Sorting Algorithms Visualizer
VisualizerInteractive visualization of classic searching + sorting algorithms for learning and demos.
Contact
For research collaborations, internships, or talks, send a short note with context + one link.
Email: bhattahan@gmail.com