open to collaborationsfreiburg · de · --:-- cetportfolio v3 · run 2026

M.Sc. Computer Science · AI · Universität Freiburg

KARANANCHAN.

I TRAIN AGENTS.

I COMPRESS MODELS.

I SHIP SYSTEMS.

I MEASURE EVERYTHING.

I work on humanoid RL, browser-based detectors, language-model systems, production RAG, and the infrastructure around long training runs. Some projects are research reproductions; others end as deployable tools. The loss curve in the corner tracks your reading progress. yes, it converges.

Start the run ↓

CV · PDF

GitHub

scroll for the projects ↓

now ⟶

still chewing on why RLPD did better with the offline data thrown out

§01

About the author

ckpt 01 · bio loaded

I'm doing an M.Sc. in Computer Science (AI) at the University of Freiburg. Before that, I completed a B.E. in Computer Science with a GPA 9.33/10.

I like the point where a paper stops and engineering begins. I rebuild results, then take care of the data, evaluations, deployment, and automation. Models are only one part of the work.

Previously: ML intern building production RAG systems at WiZdom Ed. Currently: coursework in deep learning, PGMs and robot mechanics, plus the 2026 research roadmap below. English C2 · Hindi native · German A2→B1. Off the clock: over-engineering n8n automations for my own life and defending masala chai against German filter coffee (a study with n=1 and strong priors).

Dithered duotone portrait of Karan Anchan drinking chai — fig. 0 · the authorchai, not coffee · n=1

0123456789

.33 / 10

B.E. GPA · German 1,3

0123456789

projects

2026 roadmap scope

0123456789

+ yrs

Python & PyTorch

§02

Selected work

ckpt 02 · four shipped, one in progress

Completed · 2026Reinforcement LearningLab project · team of 3

RLPD: offline-to-online RL on locomotion and Humanoid

A three-person PyTorch reproduction and critical evaluation of RLPD (Ball et al., ICML 2023). Across the complete locomotion matrix, RLPD finishes at 88–90 normalized on all three tasks. On Humanoid-v5, only 6.6% of online states are covered by the offline dataset—and online-only beats the 50/50 mix by +21.9 points at the matched 500k horizon.

88–90

minari-normalized · 3 tasks

+21.9

online-only · matched 500k

6.6%

humanoid offline-state coverage

Read the deep dive ↗Repository ↗Paper ↗

Latest diagnostic · 3 seeds

Online states covered by offline data

95th-pct NN radius

Hopper

56.2% · R 1.79×

Walker2d

69.2% · R 1.89×

HalfCheetah

71.6% · R 1.45×

Humanoid

6.6% · R 6.64×

Humanoid is the outlier: sparse coverage, 6.64× normalized distance.

fig. 1 — offline-state coverage · 3 seeds · humanoid is the 6.6% outlier

Shipped · 2026Computer VisionEdge deployment

YOLO26 at the edge: one detector, three runtimes

I fine-tuned an NMS-free YOLO26 and deployed the same network through TensorRT on an RTX 5070, ONNX Runtime on a Ryzen 7700, and WebGPU in the browser. Each path has latency and accuracy measurements; the GPU paths also include NVML power data. FP8 reaches 560 FPS, while FP16 has the best latency-per-watt result on this Blackwell GPU.

runtimes, one model

560

FPS · GPU (FP8)

FPS · in-browser

Live demo ↗Repository ↗

Accuracy cost of quantization. FP16 and FP8 stay within the 2% budget; INT8 does not.

fig. 2 · accuracy cost of quantization · measured on RTX 5070

Shipped · 2026NLP · from scratchRe-evaluated

A PyTorch Transformer for English → Hindi

A 6-layer Transformer written directly in PyTorch, without nn.Transformer or transformers. It trains on Samanantar with byte-level BPE and a Noam schedule. A new evaluation on a frozen 5k test set shows that beam search adds 0.2 chrF++ at 9.3× the latency, improving 162 sentences and worsening 140.

16.9

sacrebleu · beam k=4

41.6

chrf++ · frozen test set

~43M

params, from scratch

Repository ↗Paper ↗

Per-sentence chrF++ scatter for greedy and beam decoding. Beam improves 162 sentences and worsens 140 at 9.3 times the latency.

fig. 3 · beam search on a frozen 500-pair test

Study complete · 2026Hybrid architecturesLanguage-model systems

Mamba-2 × attention: a hybrid LM ratio study

Three 52–54M-parameter hybrid LMs interleaving Mamba-2 SSM blocks with causal attention, each trained on 700M matched OpenWebText token positions. 1:3 leads on perplexity and sampled generation; 1:15 cuts logical state by 66.3% at 8K for a 0.212 perplexity increase.

26.30

val ppl · ratio 1:3

66.3%

less state · 1:15 at 8K

52.3

tok/s · sampled generation

Showcase ↗Repository ↗

Matched 700M-token sweep

Logical inference state at 8K

lower is better

1:3

61.33 MiB
PPL 26.301

1:7

34.22 MiB
PPL 26.466

1:15

20.66 MiB
PPL 26.513

1:15 saves 66.3% vs 1:3 · +0.212 validation perplexity

fig. 4 — 8K logical state · matched 700M-token variants

In progressInterpretabilitySafety

Sparse autoencoders for tracing circuits in a small LM

Training sparse autoencoders over the residual stream of a small open LM to decompose activations into monosemantic features, then circuit-tracing induction behaviour. Early dictionaries hit ~78% auto-interp on probed layers.

16×

dictionary expansion

~78%

auto-interp score

L4–L9

layers probed

Repo coming soon ↗Circuits thread ↗

feature circuit · layers 4 → 9

fig. 5 · feature circuit, induction

§03

2026 research roadmap

ckpt 03 — a menu, not a mandate

World-model RL on Crafter

Reproduce DreamerV3 at one million environment steps, then compare imagination horizons and latent types. The evaluation will include imagined and real rollout videos.

RSSM latent dynamics + KL balancing
Actor-critic on imagined rollouts
Horizon ablation H = 5 / 15 / 30
Categorical vs Gaussian latents
Crafter score @ 1M steps

Reasoning via GRPO / RLVR

Post-train a small LM with verifiable rewards on math and measure the test-time-compute scaling curve.

server.ts

Diffusion LM from scratch

Compare a masked-diffusion language model with an autoregressive model trained at the same compute budget.

AR twin baseline on FineWeb-Edu

Masked-diffusion objective (LLaDA-style)

Matched-FLOPs eval: ppl + gen quality

Write-up: when does diffusion win?

Efficient LLM inference lab

Measure quantization, speculative decoding, and KV-cache compression on one RTX 5070, then compare throughput with retained quality.

tok/s vs fp16 baseline

87%

KV-cache saved

63%

quality retained (MMLU)

98%

§04

The record

ckpt 04 · experience and education

Oct 2023 — Oct 2024

Machine Learning InternWiZdom Ed

Built a production RAG search system with LangChain + ChromaDB over 5,000+ educational documents.
Evaluated 100 query batches: 71.7% Recall@5, 93.4% groundedness, 89.1% refusal accuracy, and 2.72 s p95 latency.

Mangalore, IN

Apr 2025 — present

M.Sc. Computer Science (AI)

Albert-Ludwigs-Universität Freiburg — deep learning, probabilistic graphical models, statistical pattern recognition, robot mechanics.

2020 — 2024

B.E. Computer Science

N.M.A.M. Institute of Technology — GPA 9.33/10 (German equivalent 1,3).

MLOps Specialization — Duke ML Specialization — Stanford / DeepLearning.AI

§05

Working stack

ckpt 05 — daily drivers first

Core

Pythondaily
PyTorchdaily
C++ / Csolid
SQLsolid
TypeScriptworking

Research

Transformers / PEFTsolid
Mamba-2 / SSMsactive
MuJoCo · Gymnasiumsolid
MONAIsolid
W&Bsolid

Systems

ONNX / TensorRTactive
Dockersolid
FastAPI / SSEsolid
CUDAlearning
Git / CI-CDdaily

Agents & data

LangChainsolid
ChromaDB / Qdrantsolid
n8nactive
MCPlearning
Ollamaworking

let's talk✳let's talk✳let's talk✳let's talk✳let's talk✳let's talk✳

ckpt 06/06 · training complete · available for interesting work

kar.anchan02@gmail.com

I'm open to research collaborations, working-student roles, and difficult engineering problems. I'm based in Freiburg and usually reply before the next training run finishes.

GitHub LinkedIn CV · PDF