How a Chinese Gym Runs 9 AI Agents on a Budget Server With No GPU

·1 views

A fitness gym in China is running nine specialized AI agents on a bare-metal Ubuntu server with just 2 CPU cores and 3.6GB of RAM, with no GPU or cloud infrastructure. The agents handle tasks ranging from member fitness report interpretation and coach scheduling to investor relations, brand content, and security auditing. Heavy AI processing is offloaded entirely to the DeepSeek API, meaning the local server only manages orchestration, sessions, and file storage. Each agent is defined by three markdown files encoding its identity, mission, and operational rules, which serve as both runtime configuration and human-readable documentation. The project demonstrates that production-grade multi-agent AI systems can be built on minimal hardware when paired with a reliable external LLM API and a well-structured orchestration layer.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Developer Series Wraps Up Full RAG System Build Using Python, pgvector, and Gemini

A multi-part developer tutorial series on DEV Community has concluded, documenting the step-by-step construction of a complete Retrieval-Augmented Generation (RAG) system from scratch using Python. The project progressed from basic database setup with pgvector on PostgreSQL through document ingestion, cosine similarity search, and a full RAG pipeline, ultimately reaching multi-step agentic loops and Model Context Protocol (MCP) server deployments. Key technical decisions included capping Gemini embeddings at 768 dimensions to comply with pgvector's HNSW index limit, and using distinct task types for document storage versus query retrieval to preserve accuracy. The free tiers of Render and Supabase were used to host the MCP server and pgvector database respectively, with a specific connection pooler port required to bridge IPv6 compatibility issues. The author noted that evaluation frameworks, observability tooling, security hardening, LLMOps practices, and fine-tuning were intentionally left out of scope for future exploration.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds Scaffold Tool to Auto-Generate Spring Boot Microservices

A developer has created a microservice generator tool called Scaffold, designed to automate the creation of new microservices. The tool is built around the Java and Spring Boot ecosystem, targeting backend developers looking to speed up project setup. The creator shared a walkthrough video demonstrating how the generator works in practice. The tool aims to improve developer productivity by reducing the repetitive boilerplate work typically involved in bootstrapping microservice projects.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Tutorial: How to Let an LLM Autonomously Decide When to Search in a RAG System

A new developer tutorial explains how to implement Tool Use in a Retrieval-Augmented Generation (RAG) pipeline, enabling a large language model to decide when and what to search rather than following a hardcoded retrieval flow. In traditional RAG setups, a search function is always called before generating an answer, but Tool Use allows the LLM to determine whether retrieval is necessary at all. The LLM is provided with descriptions of available functions and can respond with either a function call or a direct text answer based on its judgment. The tutorial uses Google's Gemini API alongside a PostgreSQL vector database, walking through a working Python implementation called 06_tool_basic.py. This approach improves response quality in cases where the user's question may already be answerable, or where multiple targeted searches with different queries would yield better results.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why pgvector, 768 Dims, and Gemini Flash: RAG Design Decisions Unpacked

A technical breakdown of a Retrieval-Augmented Generation (RAG) pipeline explains the reasoning behind key architectural choices, including using pgvector over dedicated vector databases like Pinecone or Weaviate. The author chose pgvector because it integrates with existing PostgreSQL infrastructure, supports SQL and vector search in a single query, and handles millions of documents via HNSW indexing. Google's gemini-embedding-001 model was configured to output 768 dimensions instead of the default 3072, balancing retrieval quality with pgvector's 2000-dimension HNSW limit and storage efficiency. Separate task types — RETRIEVAL_DOCUMENT for ingestion and RETRIEVAL_QUERY for querying — were used to leverage the model's asymmetric training, which improves retrieval accuracy. The HNSW index was preferred over IVFFlat for its faster query speed and higher accuracy at scale, while Gemini 2.5 Flash was selected as the answer-generation model.

0 comments Read more at DEV Community