Research themes
What I focus onInterpretable AI, transformer inference systems, failure-to-eval compilers, human-AI reliance, structured knowledge extraction, and evaluation that reveals where models actually break.
MS CSE @ UCSC
NeuroSymbolic AI
LLM Systems
Knowledge Graphs
Evaluation
Interpretable AI, backed by real systems.
I build systems that ▍
I build AI systems that pair research rigor with deployable engineering. My recent work spans transformer compiler/runtime systems, failure-to-eval compilers, human-AI reliance and observability, neuro-symbolic reasoning, multimodal safety, knowledge graph generation, anomaly detection, and time-series modeling, along with the backend, data, and cloud tooling needed to make those systems useful outside a notebook.
About
I am most energized by work that connects model behavior, system design, and clear evaluation.
Engineering stack
How I buildPython and Node.js across ML pipelines, backend services, experiment tracking, and cloud infrastructure for production-minded AI systems.
Skills
Research engineering, backend systems, and production ML.
Languages
Python, Java, C/C++, SQL, JavaScript, HTML/CSS, R
Machine Learning
PyTorch, TensorFlow, Keras, Transformers, scikit-learn, NumPy, Pandas, Matplotlib
LLM & Data Tools
LangChain, ChromaDB, RAG, vector databases, PyPDF
Backend & Deployment
Node.js, FastAPI, Flask, REST APIs, Docker, Kubernetes, Linux, Git, CI/CD, Nginx
Cloud & Data Engineering
AWS EC2, AWS S3, AWS IAM, AWS Lambda, SageMaker, CloudWatch, PostgreSQL, Redis, ETL, SQL optimization, schema design, data cleaning, time-series preprocessing
Software Engineering
TDD, API integration, version control, DSA, System Design, Performance Optimization, Code Reviews, Debugging
Experience & Education
Researcher - AI Explainability and Accountability (AIEA) Lab @ UCSC
Sep 2025 – Jun 2026
Built ReliaGuard Studio and BreakPoint for human-AI reliance detection and adversarial LLM evaluation.
- Used Next.js, TypeScript, FastAPI, PostgreSQL, Docker, and MLflow across observability and evaluation workflows.
- Combined neuro-symbolic risk modeling, conformal gating, multi-judge validation, and explainable dashboards.
- Compiled real LLM failure modes into adversarial regression suites with versioned traps and rubrics.
Data Scientist - AI Institute of South Carolina (remote)
Jun 2024 – May 2025
Researched multimodal toxicity mitigation and anomaly detection with reproducible experiment workflows.
- Used PostgreSQL-backed experiment tracking to manage dataset metadata, model outputs, evaluation metrics, and benchmarking results.
- Contributed to projects such as DE-HATE and Time Series Foundation Model evaluation.
Data Scientist - Space Applications Centre (ISRO)
Jan 2024 – May 2024
Built an end-to-end LSTM satellite clock-bias prediction platform for more reliable navigation timing forecasts.
- Added Apache Kafka live streaming and a Next.js, React, and TypeScript UI around the prediction workflow.
- Automated preprocessing and anomaly filtering to improve signal quality and downstream forecasts.
Software Engineer (AI) - Indus Institute of Technology and Engineering
May 2022 – Jan 2024
Developed Node.js-backed LLM pipelines for chat, knowledge graphs, and retrieval-augmented medical QA.
- Built a university website chatbot, automated knowledge graph generation, and medical QA workflows.
- Evaluated GPT-4, LLaMA-2, and BERT for structured semantic extraction and grounded reasoning.
University of California, Santa Cruz
Aug 2025 – Jun 2026
Master of Science in Computer Science and Engineering - Santa Cruz, California
- GPA: 3.82
Indus Institute of Technology and Engineering
Aug 2020 – Jul 2024
Bachelor of Technology in Computer Engineering - Ahmedabad, India
- GPA: 3.94
Projects
Projects highlighted in the current resume.
CacheIR
Compiler runtimeInspectable compiler and runtime for decoder-only transformer inference: imports Llama, Mistral, and Qwen-style models into a custom IR, specializes prefill/decode graphs, plans KV-cache-aware execution, selects backend kernels, and runs through a reference runtime.
BreakPoint
LLM evaluationFailure-to-eval compiler for LLM systems that mines real failure modes into versioned adversarial regression suites with DSL-defined traps and rubrics, multi-judge calibration, RAG/tool simulators, CI gates, and FastAPI/Next.js dashboards.
ReliaGuard Studio
AI observabilityNext.js, TypeScript, and FastAPI observability platform with SDKs, guardrail APIs, review queues, and conformal neuro-symbolic models to detect and audit overreliance and underreliance in human-AI workflows.
Research Forge - A Self-Improving Research Agent
Research agentUniversal self-improving AI research agent that discovers papers, generates hypotheses, runs experiments in a sandbox Python environment, and updates its reasoning strategy using long-term graph memory.
Meeting AI Assistant
Desktop copilotReal-time AI meeting copilot desktop app that captures system audio, transcribes conversations, and streams context-aware responses using the OpenAI API with screenshot and document context support.
Time Series Foundation Model Evaluation for Anomaly Detection
BenchmarkingBenchmarked TimeGPT, Chronos, and Time-MOE across multiple datasets, showing where traditional statistical and deep learning models can still outperform TSFMs.
Knowledge Graph Generation from Large Language Models
Structured extractionAutomated pipeline for generating structured knowledge graphs from unstructured text using large language models, with evaluation across GPT-4, LLaMA-2, and BERT.
Med-Bot: Retrieval Augmented Medical Assistant
RAGRetrieval-augmented medical assistant built to provide accurate and reliable answers grounded in medical PDFs.
Papers
Conference papers, workshop papers, and preprints.
Leveraging LSTM for Predictive Modeling of Satellite Clock Bias
IEEE XploreLSTM-based clock-bias forecasting with a practical preprocessing pipeline for navigation reliability.
Generating Knowledge Graphs from Large Language Models: A Comparative Study of GPT-4, LLaMA 2, and BERT
arXivStructured knowledge graph generation from unstructured text, evaluated for semantic quality and GraphRAG readiness.
Med-Bot: An AI-Powered Assistant to Provide Accurate and Reliable Medical Information
arXivRetrieval-augmented medical assistant for grounded answers from medical literature PDFs.
Time Series Foundational Models: Their Role in Anomaly Detection and Prediction
AAAI 2025 WorkshopEvaluation of time series foundation models for anomaly detection and prediction under realistic benchmarking settings.
Contact
For research collaborations, internships, or ML engineering conversations, send a short note with context and one link.
Email: bhattahan@gmail.com