OPEN TO WORK • AI / ML / GENAI / MLOPS • 2026

ML & AI
Engineer.

I am Shalin Vachheta, an M.Sc. Mechatronics student at the University of Siegen focused on building practical AI systems. My work spans LLMs, RAG, Deep Learning, and offline-first MLOps — from model development and evaluation to deployment, monitoring, and reproducible ML pipelines.

Profile

// Career Journey

DEC 2025 - PRESENT

Master Thesis

University of Siegen

MLOps Pipeline for Continuous Mental-Health Monitoring

  • Architecting a production-grade, offline-first MLOps pipeline for wearable IMU time-series (ACC/GYRO) for anxiety detection.
  • Utilizing DVC and MLflow for robust reproducibility, dataset/model versioning, and tracking artifact lineage.
  • Designing privacy-first deployments using containerized inference, FastAPI services, and structured artifact management.
MAY 2024 - MAY 2025

Student Research Assistant

University of Siegen

Sequence-to-Sequence Event Forecasting

  • Developed a BiLSTM Seq2Seq forecasting model for multivariate simulator event streams.
  • Improved performance by 25% over the baseline, gaining an additional 15% via Optuna-based hyperparameter optimization.
  • Implemented mask-aware evaluation, reliable logging, and checkpointing for processing variable-length sequences.

Featured MLOps & Systems

Vehicle Insurance Data Pipeline

  • Engineered an end-to-end ML pipeline for insurance response prediction trained on 762k+ records.
  • Achieved a 91.8% F1-score with modular pipeline stages, config-driven checks, and artifact traceability.
  • Deployed via CI/CD pipelines as a containerized FastAPI service on cloud infrastructure.
#scikit-learn#MongoDB#AWS#Docker#GitHubActions#FastAPI

GenAI, LLMs & Retrieval

RAG Chatbot — Multilingual Document Intelligence

  • Developed a multilingual RAG pipeline for scalable tender-document analysis.
  • Implemented metadata mapping, deduplication, and embedding-based retrieval.
  • Enhanced response accuracy with contextual re-ranking for enterprise workflows.
Qdrant • Ollama • Embeddings • Re-ranking • Streamlit

LLM-Based AI Email Generator

  • Built a context-aware email drafting system using Llama 3.1 and LangChain.
  • Structured persona and company inputs into reusable, semantic prompt templates.
  • Delivered a controllable generation workflow via a streamlined Streamlit interface.
LangChain • Llama 3.1 • ChromaDB • Streamlit

PEGASUS Summarization

  • Fine-tuned Google PEGASUS on 16k+ SAMSum samples for abstractive summarization.
  • Evaluated model performance using ROUGE metrics within a reproducible pipeline.
  • Packaged and deployed the final model as a modular FastAPI inference endpoint.
Hugging Face • PEGASUS • ROUGE • FastAPI • Docker

Transformer From Scratch

  • Implemented the core 'Attention Is All You Need' architecture in PyTorch.
  • Built multi-head attention, positional encoding, and residual connections.
  • Enabled sequence modeling experiments with masking and beam search integration.
PyTorch • Attention • BLEU • Beam Search

Research & Deep Learning

Seq2Seq Event Forecasting

  • Engineered a BiLSTM encoder-decoder for multivariate simulator event streams.
  • Optimized hyperparameters using Optuna to improve baseline performance by 25%.
  • Integrated mask-aware evaluation and reliable checkpointing for robust research.
PyTorch • BiLSTM • Seq2Seq • Optuna

Microscopic Denoising & Segmentation

  • Designed a dual-branch U-Net pipeline for fluorescence microscopy analysis.
  • Achieved simultaneous image denoising and precise semantic segmentation.
  • Implemented robust preprocessing, hybrid loss functions, and method benchmarking.
U-Net • OpenCV • Dice Loss • Benchmarking

Data Analytics & Classical ML

Retention Analytics — Churn Insights

  • Executed end-to-end churn analytics on 10k+ customer records.
  • Trained Random Forest classifiers reaching ~89% accuracy with feature analysis.
  • Delivered actionable, business-facing retention recommendations.
Pandas • Random Forest • EDA • Business Analytics

Predictive Modeling & Decision Support

  • Constructed a complete ML workflow spanning feature engineering to evaluation.
  • Trained and compared multiple models, achieving an R² of 0.87 with XGBoost.
  • Deployed the optimal regression model as a containerized Flask API service.
scikit-learn • XGBoost • Flask • Docker

US Vehicle Sales Forecasting & Insights

  • Analyzed 558k+ vehicle sales records with large-scale data preprocessing.
  • Engineered temporal and vehicle-level features to uncover pricing elasticity.
  • Developed 10+ EDA visualizations to highlight seasonal market behavior.
Pandas • NumPy • Matplotlib • Seaborn

IPL Match Outcome Analysis

  • Conducted exploratory analytics on 700+ professional cricket matches.
  • Studied the impact of toss strategy, venue effects, and team performance.
  • Utilized Pandas and NumPy to uncover historical outcome patterns.
Pandas • NumPy • EDA • Visualization
PythonSQLPyTorchTensorFlowscikit-learnMLflowDVCDockerFastAPILangChainTransformersRAGQdrantOllamaGitHub ActionsPythonSQLPyTorchTensorFlowscikit-learnMLflowDVCDocker