Two-Pass Grading Pipeline Uses Heuristics and Local LLM to Filter Tech Messages

A developer building a message-grading system on limited hardware — an RTX 4050 with 6GB VRAM and an Intel i7 — has detailed how the pipeline processes data without blindly feeding it to a local LLM. Before any LLM call is made, a heuristic layer using length checks and regex filters out obvious noise, eliminating roughly 12% of redundant messages in real runs. Each remaining message is then graded individually by the LLM on a scale of 1 to 5, where higher scores indicate clear technical value such as solutions, useful resources, or practical results. The system intentionally keeps heuristic filtering lenient to avoid dropping important messages, relying on the LLM as a safety net for borderline cases. Results are stored in JSON format during the MVP phase, with scalability improvements planned for later stages of development.