blog · notes from the lab
Writing & research
Long-form technical articles on ML, speech, vision, and the engineering around them. Published from plain markdown.
latest
Fine-tuning ASR models with limited GPU resources
What it actually takes to adapt a multilingual speech model to a low-resource language on one T4 — data prep, PEFT, and honest evaluation.
Read articleall articles
- VISION
Why CNNs are computationally efficient
Receptive fields, weight sharing, and the parameter count difference between dense and convolutional layers — the structural reason vision models don't blow up.
7 min readarticle 2 - SPEECH / ASR
How CTC loss works
The blank token, repeated labels, and why CTC sidesteps the alignment problem — a worked intuition for frame-level speech recognition.
9 min readarticle 3 - LLM SYSTEMS
Building document intelligence systems
Layout-aware extraction, chunking, embeddings, and retrieval — the engineering decisions that decide whether a document QA system actually answers questions.
12 min readarticle 4 - NLP / EFFICIENCY
Understanding LoRA mathematically
Why freezing a base model and training low-rank adapters works — the rank argument, the delta-weight update, and where the savings actually come from.
10 min readarticle 5 - NLP / DATA
What PCA actually does to your data
A geometric view of principal component analysis — what the eigenvectors and eigenvalues really mean, and why the 'information loss' framing is exactly right.
8 min readarticle 6