llmevaluation

An eval harness and report on some of the latest LLMs that can be run locally

Automated RAG & LLM Evaluation Pipeline — LangChain · Ragas · Jenkins · GitHub Copilot Proxy ($0 cost)

This repository is a practical overview of inference engineering for Large Language Model deployment, with vLLM as the serving engine. It is meant to help you understand not only how to start a model server, but also how to reason about throughput, latency, batching, quantization, KV cache memory, and GPU VRAM requirements before deploying a model.

AI-powered email generation assistant built with Streamlit and GPT-4o-mini or gemini, featuring prompt engineering evaluation and custom metrics comparison.

LLM prompt benchmark on 150 RAGBench samples × 5 strategies. RAG: 89.5% faithfulness, 10.5% ungrounded vs zero-shot: 31.3%, 68.7%. Qwen3 + Gemini judge.

PrefRank predicts which LLM response humans prefer using a 3-class classification framework. Combines linguistic, structural, and TF-IDF features with calibrated models (MLP, RF, XGBoost, LightGBM) — no embeddings required.

Physician-reviewed evaluation of a local medical LLM on original USMLE-style clinical questions.

Vision-language triage for forensic analysts: BLIP-2 + VQA to classify and sort images recovered from seized devices, replacing manual review of large media dumps.

A tool that scores an LLM's response to a prompt against a five-part rubric, using Claude as the judge

An agent proposes, a declarative policy authorizes, and the gate holds the credentials so the model never does. Gmail + Calendar + Notion.

Context engineering platform for AI coding agents with graph-aware retrieval, repository intelligence, MCP integration, and evaluation.

Python agent that ingests funnel data, generates a written drop-off diagnosis using Claude, and includes an A/B eval of the design decision.

On-device LLM evaluation on Apple Silicon with MLX: Laya-mlx, Llama 3.2 3B at 4-bit, 8-bit and bf16 quantization, LoRA fine-tuning and adapter fusing, compared on accuracy, speed and memory for GitHub issue classification.

A production-oriented Retrieval-Augmented Generation system with hybrid retrieval, reranking, citations, evaluation, and regression testing.

Evaluate Qwen2.5 GGUF models on Indonesian tasks with reproducible benchmarks, measuring accuracy, latency, and throughput using llama-cpp-python.