Open to AI / LLM Engineer roles · EU
Building intelligent agents & RAG systems
MSc Artificial Intelligence @ BTU Cottbus. I design LLM agents, RAG pipelines, and fine-tuned models — turning research into production-grade AI.
MSc Artificial Intelligence @ BTU Cottbus. I design LLM agents, RAG pipelines, and fine-tuned models — turning research into production-grade AI.
I'm Vidyashree — an MSc AI student at Brandenburg University of Technology, focused on LLM agents, RAG architecture, and fine-tuning.
My current work centers on bringing language models into industrial control — using LoRA/QLoRA-adapted Llama 3.2 to make hydraulic process systems adaptive and natural-language driven.
I believe in "true speed is predictable, secure, and reliable execution" — production-grade AI that works, not just demos.
Replacing rule-based PID and C++ HTTP control logic with adaptive AI decision-making. Fine-tuned Llama 3.2 3B Instruct with LoRA/QLoRA on hybrid real + physics-informed simulated data (20 sensor variables, 5s sampling). Closed-loop evaluation on level + temperature control.
Building an end-to-end agent that auto-classifies, prioritizes, and drafts replies to GitHub issues for open-source maintainers. n8n + LangGraph orchestration with RAG, evals, and Langfuse observability.
Presented 5-min thesis pitch at BTU Cottbus-Senftenberg, organized by Dr. Mahdi Taheri.
Built a tool-augmented agent on Qwen2.5-Coder-32B with dynamic tool calling, deployed as a Gradio app on Hugging Face Spaces.
Scalable RAG pipeline with semantic search over enterprise documents.
Began LLM-driven adaptive control research at BTU's Reliable & Secure Systems lab.
Started Master's program; transitioning from enterprise BI/ETL into AI engineering.
Tool-augmented agent on Qwen2.5-Coder-32B with dynamic tool calling, deployed as a Gradio app on Hugging Face Spaces.
Semantic search + LLM reasoning over unstructured enterprise documents using Chroma, RecursiveCharacterTextSplitter, and Gemini 2.5 Flash Lite.
Natural-language to SPARQL conversion with entity linking and Wikidata knowledge graph traversal.
Domain-specific Named Entity Recognition fine-tuned on industrial maintenance logs and technical documents.
Facial Emotion Recognition enhanced with Convolutional Block Attention Module for mood-aware applications.
Satellite imagery analysis using DeepLabV3 segmentation + MiDaS depth estimation for rooftop solar feasibility.
5-minute research pitch covering motivation (control gap in industrial fluid systems), approach (Llama 3.2 + LoRA/QLoRA + closed-loop control), and current results: 92.8% accuracy on temperature control, 57.8% on level control across 689 decisions.
Open to AI/ML & LLM Engineer roles in Germany or remote-friendly EU teams. Always happy to chat about agents, RAG, and production AI.