Deep dives on Machine Learning, Agentic AI, personal journeys, and lessons learned from building software.
I benchmarked a 421M open-source decision model (Laya) against a local LLM on tool routing and document typing, then wired them together. What held up, what didn't, and the silent truncation I almost missed.
I set out to build an auto-optimizing system prompt loop using local LLMs. Four rounds of debugging later, here's what I learned about GPU memory, judge consistency, overfitting, and rate limits.
What if you could run an AI model offline on your own hardware? I upgraded my RAM and set up Ollama with Qwen 2.5 Coder 3B inside PyCharm. Here is the honest truth about how it actually runs.
We're told embeddings are the only way. But what if document structure is the real key? I benchmarked PageIndex against Naive RAG across 58 tests to find out.
The unfiltered story of 7 years chasing a future I chose before I understood it.