Machine learning research and engineering

Karan Anchan

I build and evaluate models for reinforcement learning, language modeling, medical segmentation, and edge vision. The work below records methods, results, and limitations.

M.Sc. Computer Science (AI) · University of Freiburg · Freiburg, Germany

Selected work ↓GitHub ↗CV ↗

01 / Research and engineering

Selected work

Methods, measured outcomes, and limits are summarized below. Full protocols and artifacts are in the linked projects.

01Reinforcement learningTeam of three · 2026

Offline-to-online RL on locomotion and Humanoid

Reproduced RLPD in PyTorch across three MuJoCo locomotion tasks, then tested replay composition on Humanoid-v5 with matched architecture and online interaction budgets.

Tools PyTorch · Gymnasium · MuJoCo · Minari · W&B

Study and figures ↗Code and results ↗

Recorded RLPD policy rollouts for Hopper, Walker2d, and HalfCheetah shown side by side. — Recorded locomotion rollouts · qualitative, not the Humanoid ablation result.

Recorded result+22.0normalized-return points

Online-only minus 50/50 replay at the 495k-step evaluation; three seeds per condition.

Limit Descriptive difference with substantial seed variance. Minari-v5 normalization is project-specific.

02Efficient language modelsIndividual study · 2026

Mamba-2 and attention ratio study

Compared three 16-layer hybrid language models under the same tokenizer, data pipeline, optimizer schedule, and 700 million sampled training-token positions per variant.

Tools PyTorch · Mamba-2 · Hugging Face Datasets · FastAPI

Research summary ↗Code and results ↗

Animated reveal of calculated 8K logical state: 61.33, 34.22, and 20.66 MiB for the 1:3, 1:7, and 1:15 attention-to-SSM models. — Ratio-sweep figure · calculated 8K logical state, not peak VRAM.

Recorded result26.301best validation perplexity

The 1:3 attention:SSM model led the single-seed sweep. At 8K, 1:15 used 66.3% less calculated logical state for +0.212 perplexity.

Limit Matched sampled-token exposure is not matched FLOPs. Long-context retrieval remained weak in the tested setup.

03Computer vision · deploymentIndividual project · 2026

YOLO26 across GPU, CPU, and browser

Fine-tuned one YOLO26-s detector on SKU-110K and built TensorRT, ONNX Runtime, and browser WebGPU/WASM inference paths for the same dense-detection task.

Tools PyTorch · ONNX · TensorRT · ONNX Runtime · WebGPU

Live browser demo ↗Code and benchmarks ↗

YOLO26 browser sample-video detection on a store shelf with product bounding boxes. — Browser sample-video detection · separate from the TensorRT latency benchmark.

Recorded result1.9 msp50 model latency

TensorRT FP16 on an RTX 5070, with 0.06% relative mAP loss and 9.3 latency-derived FPS/W in the recorded protocol.

Limit One training run; the reciprocal rate is not end-to-end video throughput, and power was measured only for GPU paths.

04Medical image segmentationIndividual project

UNETR for 3D abdominal CT segmentation

Built a 14-label volumetric segmentation pipeline around UNETR, covering CT preprocessing, patch-based training, inference, and validation analysis.

Tools Python · PyTorch · MONAI · 3D medical imaging

Code and documentation ↗

One documented validation CT slice shown in sequence as the input image, reference segmentation, and model prediction. — One validation slice · sequential view of input, label, and prediction.

Recorded result0.8027validation Dice

Recorded on one validation split of the abdominal CT dataset.

Limit No independent test set or cross-validation result is reported.

05Sequence modelingIndividual project

English–Hindi Transformer from scratch

Implemented a six-layer encoder–decoder Transformer directly in PyTorch and compared greedy and beam decoding on a fixed evaluation set.

Tools PyTorch · byte-level BPE · beam search

Code and evaluation ↗Beam analysis figure ↗

Greedy Hindi decoding visualization with final-layer, head-averaged cross-attention highlighting the English source words. — Greedy decoding visualization from the project repository.

Recorded result41.58chrF++ with beam search

Measured on 500 frozen test pairs; beam search incurred 9.3× greedy decoding latency.

Limit The result describes this fixed test sample and training setup, not a broad translation benchmark.

02 / Project-backed practice

Tools and methods

These tools appear in the projects or documented work above. The list is not a proficiency ranking.

Modeling and training

Python, PyTorch, MONAI, Transformers, Mamba-2, Hugging Face Datasets and Tokenizers

UNETR · Mamba · translation model

Reinforcement learning

Gymnasium, MuJoCo, Minari, Weights & Biases

RLPD reproduction and Humanoid ablations

Inference and deployment

ONNX, TensorRT, ONNX Runtime, WebGPU, FastAPI, server-sent events

YOLO26 deployment · Mamba inference service

Retrieval and interfaces

LangChain, ChromaDB, React, TypeScript, Next.js

WiZdom Ed internship · project showcases · this site

03 / Experience and education

Background

Oct 2023 – Oct 2024

Machine Learning Intern · WiZdom Ed

Built and evaluated a retrieval-based study-path system over 5,000+ internal educational documents using LangChain and ChromaDB. A company-provided 100-batch evaluation recorded 71.7% Recall@5, 93.4% groundedness, and 89.1% refusal accuracy.

The recorded 2.72 s p95 is across batch-level latency values. Deployment status is not documented.

Apr 2025 – present