01Computer vision
Deepfake robustness
I test face-manipulation detectors after JPEG compression, resizing and blur, using source-video splits and uncertainty estimates.
- PyTorch
- EfficientNet
- Streamlit
Clean vs. degradation-aware training
MSc AI & RoboticsNürnberg, DE · 49.45° N
SandeepDas.
Hi, I’m Sandeep.
I turn ideas into things you can try.
Most days, that means working with images, language or data. I’m studying AI & Robotics at UTN in Nürnberg.
Take a look at my work01 / The person behind the projectsINDIA → NÜRNBERG
Curiosity usually gets me started.
Then comes the part I enjoy:
building, testing and trying again.
I moved from India to Germany to keep learning about AI and robotics. Along the way, I’ve found I enjoy the hands-on work most.
Sometimes I’m figuring out where a photo was taken. Other times, I’m working on speech or helping a robot make sense of its surroundings. I like having a question I can test.
I’m happy moving between the model and the app around it. A result becomes more useful when someone can explore it, question it and see what happens.
Outside university, I like travelling and exploring new places. I'm usually the quieter person in the group.
A little of my backgroundOn my workbench
PythonPyTorchLangChainSQLKafkaDockerReact
02 / Projects & experiments
Under the hood.
A few ideas I’ve spent time with.
Open a project to see how it works.
01
02Image geolocation
European image geolocation
A RegNet model learns to locate photos across 12 European countries, combining country, region and coordinate predictions.
- PyTorch
- RegNet
- NumPy
4.63M parameters · trained from scratch
03Geospatial ML
Nürnberg land cover
I use Sentinel-2 imagery and tabular ML to map buildings, vegetation and water, then compare the city from 2019 to 2023.
- LightGBM
- GeoPandas
- Streamlit
Five spatial validation folds
04Generative images
LoRA style lab
A team project teaching Stable Diffusion a new illustration style with LoRA and a dedicated style token.
- Diffusers
- PyTorch
- LoRA
University team project
05Tool-using assistants
Research router
Routes questions to PDFs, web search, arXiv or Wikipedia, and shows which source the assistant chose.
- LangChain
- Streamlit
- Python
PDF, web and academic source routing
06Document retrieval
Answers with sources
Combines semantic and keyword retrieval from PDFs. The answer points back to the pages that support it.
- NVIDIA NIM
- LangChain
- Python
Dense + lexical retrieval
07Natural-language analytics
Read-only SQL assistant
Asks a database a question, checks the generated SQL, and runs it through read-only access and row limits.
- SQLite
- LangChain
- Streamlit
Read-only execution and query audit
08Text summarization
Long-document summarizer
Breaks PDFs, web pages and video transcripts into paragraph chunks, then combines their summaries into useful notes.
- LangChain
- Python
- Streamlit
Paragraph-aware map-reduce pipeline
09Symbolic computation
Math, checked
SymPy solves the expression and checks the answer by substitution. The language model only explains the verified result.
- SymPy
- Python
- Streamlit
Symbolic solving before explanation
10Developer tools
Local code workbench
Five workflows for explaining, debugging, testing, refactoring and reviewing code, running locally with CodeLlama.
- CodeLlama
- Ollama
- Streamlit
Local inference · inspectable prompts
11Data engineering
Redshift query stream
Replays query metrics through Kafka, stores validated events in DuckDB, and makes the stream visible in Streamlit.
- Kafka
- DuckDB
- Docker
Optional S3 and Redshift loading
12Workload analysis
Warehouse capacity planner
Turns exported query metrics into an adjustable capacity-pressure score using queue delay, spills and concurrency.
- Pandas
- Python
- pytest
CSV input · no cloud connection needed
13Cloud application
SpeechText
Records speech in the browser, transcribes it through Google Cloud, and saves transcript metadata in Firestore.
- Node.js
- Google Cloud
- Firestore
Browser audio → stored transcript
14Full-stack application
Dataset workspace
A React workspace to upload a dataset, inspect its schema, describe an analysis goal, and review processing steps and charts.
- React
- FastAPI
- TypeScript
Upload, inspect, process and visualize
Built, tested, still learning.All repositories ↗
03 / Experience2022 — 2025
Learning
on the job.
Three internships, each giving me a different way to put AI into practice.
AI/ML Research & Development Intern
ICAR · NBSS&LUP
I built Python REST APIs and automated data-access workflows so the team could spend less time on repetitive tasks.
Python · REST APIs · AutomationNatural Language Processing Intern
ITNetworkzz
I worked on text preprocessing and evaluation pipelines, getting language data ready for model experiments.
NLP · Data preparation · EvaluationAI Intern
Coincent
I developed and evaluated deep learning models with TensorFlow and PyTorch, gaining practical experience with image classification.
TensorFlow · PyTorch · Deep learning04 / EducationINDIA → GERMANY
Where I’ve
been studying.
From my first AI courses in Nagpur to a master’s in Nürnberg.
MSc Artificial Intelligence & Robotics
Technische Universität Nürnberg
I’m studying how machines learn, see and navigate, with coursework in machine learning, computer vision, robotics and data engineering.
Expected graduation · Aug 2027BTech Artificial Intelligence
G. H. Raisoni College of Engineering, Nagpur
Grade · 8.7 / 10Higher Secondary Certificate
Saraswati Bhavan Convent High School & Junior College, Nagpur
Grade · 91.6%05 / Get in touchNÜRNBERG, GERMANY
Have somethingin mind?
I’m open to working-student roles, internships and interesting collaborations. Tell me what you’re working on.