DecodedPapers
RSS FeedAI research, decoded into working engineering practice.
I'm Mayer Antoine. DecodedPapers documents how I take architectures and techniques from recent AI/ML papers, implement them on real datasets, and adapt them to solve practical problems — using standard tools and modest compute, under the same constraints most practitioners actually face. Every post is a working implementation with documented decisions, adaptations, and tradeoffs — something you can run, customize, and carry into your own work.
Recent Posts
-
Multi-Agent Conversational Recommender System: Agentic Patterns with OpenAI Agents SDK
Build a multi-agent conversational recommender system using agentic design patterns — hub-and-spoke orchestration, parallel agent execution, and reflection loops with the OpenAI Agents SDK.
-
Know Your RAG — Building an Evaluation Dataset for Public Health RAG Systems
Build a domain-specific RAG evaluation dataset from public health literature using the Know Your RAG framework — LLM-generated questions, critique filtering, and RAGAS benchmarking of Agentic vs Naive RAG pipelines.
-
Multi-Document Summarization with RAG: Hybrid Retrieval, Debate Prompting, and SPECTER
Implement multi-document summarization using RAG and hybrid retrieval (ChromaDB + ColBERT) — debate prompting, SPECTER embeddings, and working Python code based on the X-Sum framework.
-
Build an AI Literature Review Tool with RAG and LLM Debate Ranking
Build a cost-effective AI literature review tool using RAG, BM25 hybrid retrieval, and debate-style LLM ranking — based on the LitLLM paper, with a working CLI and web app.
-
Evaluating Agentic RAG Systems - A Practical Guide with PubMedQA
Practical guide to evaluating agentic RAG on PubMedQA using RAGAS — covering context precision, faithfulness, answer correctness, and exact match with ChromaDB vector store.
-
Evidence-First AI - Implementing Agentic RAG with OpenAI Agents SDK
Updated:Build a PaperQA-inspired agentic RAG system using OpenAI Agents SDK — multi-agent orchestration, semantic search with ChromaDB, ReAct pattern, and LLM-as-judge reranking.