Stockholm, Sweden

Vaibhav
Saxena

I lead Machine Learning projects by day and ping open LLMs by night. Ten years of engineering; the last several spent deep in open-source LLMs — pretraining, post-training, inference — and in building state-of-the-art solutions around them — at work and on my own time.

Lead Machine Learning Founder · barpath.ai Powerlifter · 7 yrs coached
See what I build GitHub
Vaibhav Saxena in a wine cellar, raising a glass
somewhere below sea level, among the bottles
01

Who I am

I'm an ML engineer who likes the whole stack of a problem — from tokenizer to product. Professionally I lead machine-learning projects in Stockholm, working on document intelligence and language models for the Nordics. Personally, I ship things: a coaching platform for lifters built on computer vision and open LLMs, a sunshine map of Stockholm, research on very small language models, and a steady stream of paper reimplementations because reading is never quite enough.

Open-source LLMspretraining · fine-tuning · inference · data sovereignty
Document Intelligencefrom-scratch Nordic models, medical-domain OCR & retrieval
Under the barseven years of coached powerlifting — the reason barpath exists
02

Things I've shipped

Live products, built end-to-end — design, models, infra, the lot.

Flagship Live

barpath.ai

AI biomechanics for powerlifting — point your phone at the bar and lift.

A full coaching platform built on an in-house computer-vision and AI-coaching stack. Film a set; barpath reads it like a coach with instruments.

  • Bar-path tracking, pose estimation & rep detection — velocity, fatigue, sticking points, tempo
  • Velocity-based RPE that calibrates to each lifter from their own ratings and bar speed
  • Every lift phase rebuilt as a rotatable 3D body with true joint-angle tables
  • Technique verdicts measured, not guessed — squat depth called in centimeters
  • Spot, an AI coach that knows your program, your numbers and your food
  • Adaptive periodized programs + set logger with e1RM trends
  • Barpath Kitchen: meals planned around training days — with macro-identical cuisine remixes grounded in flavor-network research (Ahn et al., 2011)
  • Full try-before-signup demo — no account needed
barpath.ai
LiveStockholm

Olisol

Which terrace is in the sun, right now?

A live sunlight map of Stockholm. Every café, terrace and park spot, marked sunny or shaded for any minute of the day — drag the time slider and watch the city's light move. Never sit in the shade by accident again.

Open Olisol
Olisol — a map of Stockholm showing sunny and shaded spots with a time slider
03

Open source & research

Paper reimplementations, from-scratch Nordic pretraining, and a complete Scandinavian medical-NLP pipeline — OCR to retrieval to models.

Paper implementations

refrag

★ 17

Implementation of Meta's REFRAG (arXiv:2509.01092): low-latency RAG decoding that compresses ~80% of retrieved chunks into embeddings with a selective-expansion policy — up to 30× faster time-to-first-token and 16× longer effective context.

PyTorchDistilBERTRAG

VectorLM

Continuous Autoregressive Language Models (CALM): a VAE folds K tokens into one latent vector, and a transformer with an energy-based head predicts the next vector — K× fewer generation steps, evaluated with the likelihood-free BrierLM metric.

PyTorchVAEEnergy head

Nordic pretraining, from scratch

MaveraLM-pretrain

MaveraLM-4B — a 4-billion-parameter transformer for Swedish, Danish and Norwegian, written from scratch: 32 layers, GQA, SwiGLU, RoPE, 12k context, pretrained on 50B tokens of Nordic FinePDFs.

PyTorch4B paramsGQA · SwiGLU

scandic-gemma3 · package

Gemma-3 architecture rebuilt in full — sliding-window attention, dual-base RoPE, GQA — with a custom Scandinavian tokenizer and streaming pipeline over mC4, ScandiQA and the Nordic Pile. Configs from 340M to 1B.

PyTorchGemma-3 archmC4 · Nordic Pile

scandic-gemma3 · script

The single-file sibling: one script that trains its own BPE tokenizer, streams five Scandinavian corpora, and pretrains a ~1B Gemma-3-style model with mixed precision, cosine LR and atomic checkpointing.

PyTorchBPE1B config

finetune-gptoss_scandic

GPT-OSS 20B taught to write Scandinavian medical journals: 4-bit quantized with LoRA on ~0.04% of weights, tuned on curated Norwegian corpus texts plus synthetic tri-lingual doctor conversations — 15 minutes on one A6000.

UnslothLoRA r=16GPT-OSS 20B

The medical-NLP pipeline

kosmos_ocr_database

Batch OCR at corpus scale: walks directory trees of PDFs, renders pages, and reads them with Kosmos-2.5 on GPU — the pipeline behind a 50,000+ page Swedish medical corpus.

Kosmos-2.5pdf2imageCUDA

finetune_kosmos_ocr

Grounded OCR training for Kosmos-2.5 on DocLayNet — text plus bounding boxes, with accelerate-based training and inference that draws its predictions back onto the page.

Kosmos-2.5DocLayNetaccelerate

ocr_evaluation

The measurement layer: a FastAPI bench that pits Tesseract, Nanonets and EasyOCR against ground truth with character/word accuracy, Levenshtein, BLEU and a composite score.

FastAPILevenshtein · BLEU

hybrid-search

RAG-ready retrieval over multilingual medical documents: Vespa hybrid BM25 + dense search with multilingual MiniLM embeddings, sub-100ms queries, and a source-attributed /rag endpoint.

VespaMiniLMFastAPI

micro-llm

The research capstone: sub-2M-parameter BERT-style micromodels with an 8k Swedish-medical vocabulary, trained on the OCR corpus — clinical NLP that runs anywhere, GDPR-first.

1.8M paramsCustom tokenizerEdge

Utilities & mischief

pdf_file_server

A single-file localhost server for browsing and reading PDF collections across many directories — the quiet workhorse of the OCR projects.

PythonZero deps

take_a_nap

Park your cursor in a glowing circle, get told "sweet dreams", and a Flask + PyAutoGUI daemon keeps your status green while you rest. Everyone deserves a nap.

FlaskPyAutoGUI
04

Miscellaneous

Smaller things that live on the internet because I wanted them to exist.

LITM Research

An interactive study of the lost-in-the-middle effect — how long-context LLMs forget what's buried mid-prompt.

litm-research

Repo Flattener

Turns an entire repository into one clean, LLM-ready file — context windows love it.

repo-flattener

Micromodels Paper

Published research: sub-2M-parameter language models that read Swedish medical OCR text — privacy by design.

read the paper

Meal Prep

A meal program built for two — shared weekly plans for couples who train and cook together.

mealprep

Workout Tracker

A no-friction training log — sets, loads and progress without the bloat of a fitness app.

workout-tracker

Cut Plan

My own cutting protocol as a living dashboard — targets, training-day cycling, the works.

cut-plan
05

Ask my personal LLM

Questions about Vaibhav? An open-weights qwen3.6 answers them — running on a GPU tucked inside the TV bench in his living room. If it's slow, cut it some slack: great furniture, limited airflow.

vaibhav-bot · qwen3.6 · tv-bench, stockholm
Hej! I'm running inside Vaibhav's old computer in his living room — inside of a freaking TV bench — and I know a suspicious amount about him. Work, projects, lifting numbers — ask away.