tags
Click a filter to scope results to what I've built or knowledge hub.
I used Datalab's Surya OCR 2 from Hugging Face to process an 800+ page Hindi and Sanskrit book, preserving text, formulas, tables, reading order, and even vertical table layouts with a cache-first pipeline.
A 3B vision-language model from Baidu that parses whole documents — single images, multi-page files, and PDFs — in a single inference pass, with bounding boxes and formatting preserved.
One dotnet new command gives you a Blazor chat-with-your-data app — vector search, document ingestion, citations, and Azure OpenAI wired up on keyless auth.
Where does the conversation history live? That one decision shapes your agent's cost, privacy, portability, and UX. Here's how service-managed and client-managed storage differ — and how Microsoft Agent Framework abstracts both behind one programming model.
A .NET console chat app in one Program.cs file — using Microsoft.Extensions.AI so the OpenAI model behind it is a swappable detail, not a lock-in.
Run a model on your own machine during development, swap to Azure OpenAI in production — same code. Here's how Aspire and Ollama make local AI a registration detail, not a rewrite.
Four libraries cover almost everything you'll do with AI in .NET. Here's how Microsoft.Extensions.AI, VectorData, the Agent Framework, and MCP fit together — without juggling vendor SDKs.
Zed is a GPU-rendered code editor built in Rust by the Atom creators — millisecond startup, real-time collaboration, and built-in AI out of the box. Here's my take.
Your AI outputs are only as good as your inputs. Here's how I use prompt engineering techniques to get consistently strong results from ChatGPT, Copilot, and image generators.
A practical workflow and tool setup for generating relevant, high-intent image keywords using Everypixel, helping creators and stock contributors improve discoverability and metadata quality.