Mehul Mittal - AI/ML Engineer

About Me

I work on retrieval and evaluation for clinical trial matching.

My current research is TAIM, a benchmark that runs published trial-matching systems under a single fixed protocol so their results can be compared. Most of the work sits in retrieval and in the measurement itself: what was scored, on which corpus, and whether a run can be rebuilt afterwards. Joint work with a collaborator at TU Munich, on NSF ACCESS allocations (Delta, Expanse). I am extending it toward trial-to-patient retrieval, where the query is a study and the collection is a hospital’s own patients. Emtrial is the clinical side of that work.

Before this I spent four years building ML in production — agentic LLM pipelines, computer vision for industrial safety, and the serving infrastructure behind them. That is still how I work: versioned data, tracked experiments, and results that rebuild from their artefacts.

M.Sc. in Artificial Intelligence, FAU Erlangen-Nürnberg, 2026, with a graduation scholarship. Based in Fürth, Bavaria. Valid job-seeker permit, no visa sponsorship required.

4+ Years of Experience

€2M Business Impact

10+ Projects Shipped

375k Trial corpus benchmarked

Experience

AI Engineer

Robert Bosch GmbH

Mar 2025 – Mar 2026

Deployed production GenAI applications including RAG pipelines and multi-agent orchestration frameworks, cutting manual document review cycles by 60% and API costs by 35%. Built LLM evaluation and monitoring systems across containerized deployments.

LangGraph LangChain GPT-4o RAG Multi-Agent PyTorch AWS

AI Researcher

MaD Lab, FAU Erlangen-Nürnberg

Mar 2024 – Apr 2025

Developed deep learning models on multimodal time-series and biomarker data, improving cross-subject generalisation F1 by 12% via transfer learning and self-supervised pre-training. Co-authored a paper at IEEE-EMBS BHI 2025.

Transfer Learning Self-Supervised Learning Signal Processing PyTorch Azure ML Biomedical Sensors

Lead ML Engineer

Concentrix

Oct 2021 – Feb 2023

Architected and maintained a production ML platform serving 200+ users. Managed the full MLOps lifecycle — from data pipeline orchestration to continuous deployment — delivering €2M ROI through automated forecasting and anomaly detection.

MLOps XGBoost Explainable AI Docker & Kubernetes CI/CD AWS

ML Engineering Intern

Mirrag AI

Sep 2021 – Nov 2021

Built and evaluated ML models for early-stage AI product features, working on data pipelines and model prototyping in a fast-paced startup environment.

TensorFlow Computer Vision Object Detection Azure Python Data Pipelines

Recommendations

What colleagues and managers say

Mehul joined my team at Bosch and quickly became a valuable contributor. I was consistently impressed by his technical foundation in AI and Data Science — he shows a strong grasp of the ML pipeline and a great ability to translate complex assignments into clean, functional code. He is a fast learner who takes ownership of his work, and I recommend him to any team looking for a dedicated and capable AI Engineer or Data Scientist.

What stood out the most was how effectively Mehul streamlined forecasting workflows and turned complex data into practical insights that significantly improved forecasting accuracy. His approach was always structured, data-driven, and focused on solving real business problems rather than just building models. I would highly recommend him to any organisation looking for someone with strong analytical skills and the ability to drive impactful process improvements.

Key Projects

★ Current Research

TAIM — A Benchmark for
Clinical Trial Matching

2026 – present · with TU Munich · NSF ACCESS (Delta, Expanse)

Evaluation harness running published trial-matching systems and dense/sparse retrieval baselines over TREC Clinical Trials 2021–2023 and SIGIR CT 2016 under one fixed protocol.
Adapters and run contracts that pin each run’s corpus, model revision, configuration, retrieval depth and output hashes, so any result can be reconstructed from its artefacts.
Two-arm retrieval panel (BM25 plus a dense encoder) with cross-encoder reranking; audits of judgment coverage and input truncation across the 375,580-trial corpus.
Extending toward trial-to-patient retrieval: the query is a study, the collection is a hospital’s patient records.

Information Retrieval Dense & Sparse Retrieval Cross-Encoder Reranking TREC Evaluation Reproducibility ACCESS HPC

375,580 Trials in the evaluation corpus

4 TREC / SIGIR datasets

2.1M Eligibility criteria indexed

HPC NSF ACCESS GPU clusters

★ Master's Thesis

Transfer Learning for
Biomedical Signals

October 2025 – March 2026 · FAU Erlangen-Nürnberg

Trained Transformer and CNN models on 4 TB of EEG brain data on an NSF-funded ACCESS GPU cluster, benchmarking EEGNet, EEGConformer, ATCNet, and Vision Transformers.
Engineered chunked streaming, prefetching, and on-the-fly augmentation pipelines to maximise GPU utilisation across large-scale distributed training jobs.
Coordinated NCCL communication and batch scheduling across multiple cluster nodes; evaluated cross-subject generalisation across 6 cognitive tasks spanning passive and active paradigms.
Deployed reproducible research infrastructure with MLflow experiment tracking, model registry, and DVC dataset versioning.

PyTorch Lightning Transformers Distributed Training NCCL Domain Adaptation MLflow DVC ACCESS HPC

4 TB EEG training data processed

6 Cognitive tasks evaluated

4 Architectures benchmarked

HPC NSF ACCESS GPU cluster

Academic Capstone

Healthcare Analytics with LLMs

Oct 2024 – Apr 2025

Built a production GenAI pipeline integrating GPT-4 and Claude with RAG architecture, tool-calling agents, and semantic vector search for domain-specific medical knowledge retrieval. Designed an LLM observability framework with hallucination scoring, drift detection, and MLflow 3.0 governance meeting healthcare compliance standards.

GPT-4 Claude RAG Pinecone ChromaDB MLflow

Published · IEEE-EMBS BHI 2025

Subjective Motor Activity & Gait in Parkinson's Disease

Mar 2024 – Apr 2025 · MaD Lab, FAU

Statistically evaluated 15 stride-level gait parameters derived from foot-worn IMUs against patient-reported motor activity levels across 33 PD patients over 186 days in free-living conditions. Showed that subjective daily reports mirror objective gait changes at both macro (stride count) and micro-parameter level — providing groundwork for integrating PROs into digital PD monitoring systems.

IMU Sensors Gait Analysis Statistical Analysis Azure ML IEEE-EMBS BHI 2025

Academic Milestone

Bradykinesia Detection from Wearable Sensors

Mar 2024 – Apr 2025 · MaD Lab, FAU

Developed deep learning models on multimodal time-series data from foot-worn IMUs to detect bradykinesia in Parkinson's patients. Applied transfer learning and self-supervised pre-training techniques to improve generalisation across patients in free-living conditions.

PyTorch Transfer Learning Time-Series Wearable Sensors

More on GitHub

Computer vision, NLP pipelines, MLOps tooling, and more — all with detailed documentation and production-ready code.

View GitHub

Technical Skills

GenAI & LLMs

GPT-4o Claude API Llama 3 Gemini Hugging Face LoRA / QLoRA Prompt Engineering Tool Calling Structured Outputs

Agentic & RAG

LangGraph LangChain LlamaIndex RAG Pipelines Semantic Search Pinecone ChromaDB FAISS MCP

Deep Learning

PyTorch PyTorch Lightning TensorFlow / Keras Transformers Vision Transformers CNNs / RNNs Transfer Learning Self-Supervised Learning

ML & Data Science

Scikit-learn XGBoost LightGBM Time-Series Forecasting Anomaly Detection Feature Engineering SHAP / LIME Domain Adaptation

MLOps & Infra

MLflow Weights & Biases DVC Docker Kubernetes GitHub Actions FastAPI LLM Evaluation Drift Detection

Cloud Platforms

AWS SageMaker AWS OpenSearch Azure ML Databricks Lambda EC2 / S3 Google Cloud AI

Languages & Tools

Python SQL C++ JavaScript Bash REST APIs Git / GitHub Jupyter ONNX / TensorRT

Specialised Areas

Computer Vision NLP Biomedical AI Wearable Sensors EEG / Time-Series HPC / Distributed Training Model Quantization

Software Engineering

System Design Microservices REST APIs Event-Driven Architecture SQL / NoSQL PostgreSQL Redis Message Queues CI/CD OOP / SOLID Design Patterns Unit Testing

Languages

English — Fluent German — B1 Hindi — Native