writing
-
Evaluating metadata classification for a retrieval pipeline
Retrieval was tuned, so I measured the next stage: how reliably a local model labels each document. Across input windows, more than half the document types stayed wrong, and the chunking setting turned out to be quietly controlling what the classifier could see.
-
Evaluating metadata classification
Retrieval is tuned. The limiting factor now is metadata classification — the local model that labels each document is wrong more often than it's right. Here's the problem in detail, and how I plan to measure and improve it.
-
Evaluating document chunking for retrieval pipeline
How I evaluated the chunking layer of FindFiles by sweeping chunk size, overlap, and splitting strategy against a labeled set of 53 real government forms, on-device.
-
Semantic File Search with LangChain and Chroma
A walkthrough of a semantic file search assistant that indexes local documents and enables intelligent retrieval using vector embeddings.
-
Apache Flink
Explores how Apache Flink powers real-time fraud detection in financial systems through dynamic model updates, stateful stream processing, and scalable architecture.
-
Perceptron to MNN
A walkthrough from perceptrons to multilayer neural networks, explaining linear models, equations, and how input data is structured for deep learning.
-
Why deep learning?
The purpose of this article is to understand why one might choose deep learning algorithms over the machine learning algorithms.
-
Walls and Gates — From OGM to OM
A follow-up walkthrough simplifying the original walls-and-gates solution.
-
Walls and Gates — Breadth-First Search
Using BFS to compute the distance from every empty room to its nearest gate.
-
Validate Subsequence
A simple framework for approaching array problems — what to try first.
-
Three Number Sum
Another array problem demonstrating tradeoffs between brute force and pointer-based solutions.
-
Two Number Sum
A classic array problem that surfaces the time-vs-space complexity tradeoff.
-
Traversing a Matrix — Flood Fill
Matrix traversal with neighbor validation — the classic flood-fill pattern.
-
Kth Largest Value in a Binary Search Tree
Using in-order traversal of a BST to pull out the kth largest value.
-
Binary Tree to a Linked List
Flattening a binary tree into a linked list — combining tree traversal and linked-list manipulation.
-
Find the Town Judge
Walking through the UMPIRE method to identify the judge from trust relationships.
-
Choosing the loops!
A short article explaining how one can choose what loop to use during programming.
-
My first blog post
A short test post — the start of this blog.