M.Sc. CS · AI · Open to collaborate

Karan Anchan.

Building end-to-end AI systems — from volumetric computer vision systems, to NLP systems, to RAG pipelines in production.

Generative AI / Computer Vision / NLP

Read the work Curriculum vitae Correspond

Turn the page

§ 02 About

Building end-to-end AI systems that solve real-world problems.

I'm a student pursuing my M.Sc. in Computer Science at the University of Freiburg, with an interest in building systems for generative AI, computer vision, natural-language processing and reinforcement learning.

My work sits at the intersection of research and engineering — from volumetric image segmentation with vision transformers, to deploying retrieval-augmented generation pipelines in production, to monitoring and maintaining built systems.

I prefer building measured metrics, reproducible pipelines, and systems that ship.

Languages5

Python C / C++ SQL JavaScript React JS

AI · ML9

PyTorch TensorFlow Hugging Face MONAI LangChain Gradio ScikitLearn MLflow ONNX

Infrastructure4

Docker Kubernetes DVC AWS Sagemaker Prometheus Git CI / CD

Spoken3

EnglishC2 HindiNative GermanA2→B1

§ 03 Selected work

Studies in vision, language, and medical AI.

Case i.

Computer Vision · Medical Imaging

UNETR — 3D Abdomen Segmentation

A volumetric segmentation pipeline using the UNETR architecture — a vision transformer encoder paired with a CNN decoder — to identify 13 distinct abdominal organs. Used HU-windowing and isometric resampling to normalize soft-tissue density across the BTCV dataset.

0.82Mean Dice

13Organs

3DVolumetric

PyTorchMONAIViT

View on GitHub

Fig. UNETR BTCV / 13 organs

cross-sectional anatomy representing UNETR 3D segmentation

Case ii.

NLP · Low-resource Translation

Neural Machine Translation — English → Hindi

A 43M-parameter Transformer encoder-decoder built from scratch in PyTorch — no nn.Transformer, no high-level libraries — faithfully reproducing "Attention Is All You Need". Trained on 500k pairs of AI4Bharat's Samanantar corpus with byte-level BPE, a Noam schedule, and length-normalized beam search; ships with a Gradio demo that visualizes cross-attention live.

16.93SacreBLEU

41.58chrF++

43MParams

PyTorchHugging FaceGradio

View on GitHub

Fig. Decoder self-attention EN → HI

Decoder self-attention heatmap from the English-to-Hindi Transformer, showing the lower-triangular causal pattern

§ 04 Experience

Shipping AI in production environments.

Oct 2023 — Oct 2024 Mangalore, India

Machine Learning Intern

WiZdom Ed

Built a Retrieval-Augmented Generation pipeline using LangChain and ChromaDB for internal document retrieval, with a cosine-similarity feedback loop monitoring retrieval quality and answer relevance over time.

40%

Faster data ingestion via recursive text splitting

90%

Answer accuracy through cosine-similarity feedback loop

5,000+

Documents processed for automated student support