Notes on ML engineering, attention research, and shipping AI systems that work on real hardware.

• September 20, 2026 · 11 min

  A³ Attention: The Softmax Is Not Sacred

  Transformers have used the same fixed softmax since 2017. A³ asks a simple question: what if attention could adapt its own activation?

  • Research
  • Transformers
  • Attention

• August 5, 2026 · 9 min

  Shipping ML Products, Not Just Models

  A model that only lives in a notebook helps no one. Lessons from building BurnoutCompass AI end to end — from ensemble training to a deployed FastAPI backend.

  • MLOps
  • FastAPI
  • Lessons

• June 12, 2026 · 10 min

  LLMs on a CPU: QLoRA, GGUF and llama.cpp

  No GPU? No problem. How I fine-tune small models with QLoRA and serve them on plain CPUs with quantization and llama.cpp.

  • LLM
  • Quantization
  • llama.cpp

• May 28, 2026 · 9 min

  RAG Is a Retrieval Problem Wearing an LLM Costume

  I've built the same RAG loop three times now. Every time the failure was retrieval, not the model — here's what I changed after each one.

  • RAG
  • LangChain
  • LLM

• April 14, 2026 · 8 min

  Embeddings Beat Ratings: Rebuilding a Movie Recommender

  Instead of a user × item rating matrix, Movie-Recommender learns a vector space with a Siamese Transformer and searches it with FAISS.

  • Recommendation
  • Transformers
  • FAISS

• March 9, 2026 · 8 min

  Human Activity Recognition From Sensors, Not Cameras

  A CNN-LSTM on phone accelerometer and gyroscope streams — and why the experiment tracker ended up mattering more than the extra layer.

  • Deep Learning
  • PyTorch
  • MLflow

• February 18, 2026 · 6 min

  A Face Detector in a Weekend (and Why It's Still FastAPI)

  OpenCV, a Haar Cascade and a thin FastAPI service. What a deliberately small computer-vision project taught me about picking boring tools.

  • Computer Vision
  • OpenCV
  • FastAPI

• January 22, 2026 · 8 min

  Getting Structure Out of an LLM

  Resume-Screening-System is a RAG pipeline with a strict output contract. Here's how I keep a local Mistral model honest over hundreds of CVs.

  • LLM
  • LangChain
  • Mistral

• December 10, 2025 · 6 min

  A CRUD API That Stays Boring

  Smart Task Manager has no machine learning in it at all — FastAPI, SQLAlchemy, SQLite and a Streamlit UI. It's still the cleanest thing I ship.

  • FastAPI
  • SQLAlchemy
  • Backend