Initializing neural pathways

AI Research Intern @ ABB · Previously DFKI

Vishal Banwari

Human in the loop. Code in the wild.

MSc Computer Science @ RPTU Kaiserslautern
Specializing in Intelligent Systems & Scientific Computing

Scroll to explore

My Neural Network

Each node represents a project or skill. Connections show technologies used together.
Click any node to explore details. Watch the network think.

Project
Skill
Domain

Experience Timeline

Neural pathways forged through industry and research.

Jan 2026 – Present

AI Research Intern

🤖 ABB · Kaiserslautern, Germany
  • Building root cause analysis for logs using LLMs
  • Developing RAG pipelines and Streamlit interfaces for log intelligence
  • Applying LLM-based reasoning to industrial fault diagnosis workflows
LLMs RAG Streamlit Log Analysis
Apr 2024 – Dec 2025

Research Assistant

🧪 DFKI · Kaiserslautern, Germany
  • Developed Python GUI tools for human-in-the-loop ML workflows
  • Created novel loss functions with real-time feedback signals
  • Built visualizations for model confidence & embedding insights
  • LLMs and AI Agents energy consumption & efficiency benchmarking
Python PyTorch Human-AI Visualization
Jun 2021 – Feb 2023

Senior Engineer

🔧 Rugged Monitoring · Hyderabad, India
  • Full-stack product development using Angular + .NET Core
  • Designed and delivered industrial monitoring dashboards
  • Led backend architecture for smart data analytics features
Angular .NET Core Full-Stack Analytics
Feb 2017 – May 2021

Software Engineer

🖥 Qualitrol · Ahmedabad, India
  • Built WPF desktop applications using C++/C#
  • Integrated ML modules for energy sector systems
  • Enhanced UI/UX for electrical data visualization tools
C++ C# WPF ML Integration

Publications

2025

This Is Taking Too Long — Investigating Time as a Proxy for Energy Consumption of LLMs

We investigate inference time measurements as a proxy to approximate the associated energy costs of API-based LLMs, comparing estimations with actual measurements from locally hosted equivalents. Shows that time measurements allow us to infer GPU models for API-based LLMs, grounding energy cost estimations for end users.

Inference time as energy proxy 🔍 GPU model inference 🌿 API cost transparency
📄 Read on arXiv
AAAI 2025

Promoting Sustainable Web Agents: Benchmarking and Estimating Energy Consumption through Empirical and Theoretical Analysis

An exploration of the energy and CO₂ costs of web agents (e.g. OpenAI Operator, Google Mariner) from empirical and theoretical perspectives. Shows how different agent design philosophies severely impact energy use — and that more energy consumed does not necessarily equate to better results.

🌿 Web agent sustainability 📊 Energy & CO₂ benchmarking 🔎 Transparency in agent design
📄 Read on arXiv
HHAI 2025 · May 2025

Human in the Latent Loop (HILL): Interactively Guiding Model Training Through Human Intuition

We present HILL, an interactive framework allowing users to incorporate human intuition into the model training loop by reshaping latent space representations. The modifications are infused into training via a novel distillation-inspired approach, treating the user's reshaped latent space as a teacher.

🧠 Interactive latent guidance 🎓 Distillation from user-shaped space 📈 Improved generalization
📄 Read on arXiv

Tech Stack

The synapses that power my work.

💬

Languages

Python C++ C# JavaScript SQL HTML/CSS
🧪

ML / AI & NLP

PyTorch TensorFlow Hugging Face LangChain scikit-learn OpenCV LoRA / PEFT RAG Sentence Transformers
🌐

Web & Tools

Angular .NET Core Streamlit Gradio Git Docker Linux WPF
🚀

Focus Areas

Generative AI NLP Computer Vision Deep Learning Human-AI Collaboration Continual Learning RAG Systems AI Sustainability

Beyond Code

🎸Guitar
🏓Table Tennis
📚Reading
🏃Running
💪Gym
⛰️Hiking
♟️Chess
🎹Piano
🌱Gardening

Let's Build Something Together

Whether it's a research collaboration, engineering challenge, or just a conversation about AI — I'd love to hear from you.