FROM THEDESK
Notes on the things I actually build — machine learning, computer vision, and GenAI agents. Fewer buzzwords, more of what worked and what didn't.
- Machine Learning
How gradient-boosted trees actually work
XGBoost feels like cheating until the day it does something weird. Here is the from-scratch mental model — residuals, similarity scores, gain, and why the learning rate is really a humility knob.
Jun 14, 2026 · 10 min read
- Machine Learning
Graph-based anomaly detection for time series
Real systems are never one wiggly line. When dozens of series move as a system, the interesting failures are broken relationships — and that is exactly what a graph of time series is built to see.
Jun 2, 2026 · 11 min read
- Machine Learning
XGBoost vs LightGBM: which gradient booster should you reach for?
They both win Kaggle competitions, but they grow trees in fundamentally different ways. A practical, from-experience breakdown of when leaf-wise beats level-wise — and when it quietly overfits you.
May 18, 2026 · 9 min read
- Computer Vision
Building Rakshak: real-time crowd safety with YOLOv8
Counting 500 people in a frame is the easy part. Turning those boxes into a trustworthy NORMAL / WARNING / CRITICAL signal — before a stampede — is where the real engineering lives.
Feb 2, 2026 · 11 min read
- GenAI
A zero-hallucination RAG broker on a 100% free stack
How I wired BeautifulSoup, FAISS, and a small LLM into a property assistant that never invents listings — and then let it book its own site visits through the Google Calendar and Gmail APIs.
Mar 20, 2026 · 10 min read