ai-engineering-hub

Routes natural language queries to either vector semantic search or SQL database queries using Cleanlab Codex for intelligent decision-making. Implements a dual-path retrieval system where incoming queries are analyzed to determine optimal data source (unstructured documents via vector embeddings or structured data via SQL), then executes the appropriate retrieval pipeline and merges results. Uses LlamaIndex as the orchestration layer with Milvus or Qdrant for vector storage and SQL connectors for database access.

Enables semantic search over code repositories by parsing source code into syntax-aware chunks using tree-sitter AST parsing, then embedding and indexing these chunks with structural context preserved. Implements code-specific retrieval that understands function boundaries, class hierarchies, and import relationships rather than treating code as plain text. Integrates with LlamaIndex for embedding and vector storage, with custom chunking strategies that respect code structure and maintain semantic coherence across function/class boundaries.

Implements conversational systems with persistent memory using Zep or similar memory management systems that store conversation history, user context, and extracted facts across sessions. Maintains conversation state including user preferences, previous questions, and domain-specific context. Integrates with chat interfaces (Chainlit) to provide multi-turn conversations where agents can reference previous interactions. Supports memory summarization to manage token limits while preserving important context.

Provides MCP server implementation for audio analysis tasks including speech-to-text transcription, speaker diarization, emotion detection, and audio classification. Integrates AssemblyAI for transcription and diarization, with custom models for emotion and classification tasks. Exposes audio analysis capabilities through MCP protocol for standardized access across different clients. Supports streaming audio processing for real-time analysis.

Integrates Pixeltable (a multimodal data management system) through MCP protocol to enable structured management of images, videos, and other multimodal data alongside metadata and computed features. Provides MCP server that exposes Pixeltable operations (data ingestion, feature computation, querying) to LLM clients. Enables agents to manage and query multimodal datasets without direct database access, with automatic feature computation and versioning.

Implements a multi-agent system (via CrewAI) for content creation workflows where specialized agents (planner, writer, editor, reviewer) coordinate to produce high-quality content. Agents have specific roles with defined tasks and can iterate on content based on feedback. Supports content planning, drafting, editing, and quality review in a coordinated workflow. Integrates with RAG for research and fact-checking during content creation.

Implements a specialized multi-agent system for documentation and research workflows where agents (researcher, analyst, writer) gather information, analyze findings, and synthesize documentation. Agents coordinate to research topics, extract key insights, and produce comprehensive documentation with citations. Integrates with RAG for document retrieval and web browsing for current information. Supports automated generation of technical documentation, research reports, and knowledge bases.

Implements a specialized multi-agent system for travel planning and booking where agents (planner, researcher, booker) coordinate to gather travel requirements, research options, and execute bookings. Agents have access to travel APIs (flights, hotels, activities) and coordinate to create comprehensive travel itineraries. Supports multi-step workflows including destination research, option comparison, and booking confirmation. Integrates with external travel services through tool integration.

Implements a feedback loop that evaluates retrieval quality after initial document retrieval and automatically triggers corrective actions (re-ranking, query reformulation, or expanded search) if confidence scores fall below thresholds. Uses LLM-based relevance scoring to assess whether retrieved documents actually answer the query, then applies corrective strategies: query expansion, semantic reformulation, or fallback to broader search parameters. Integrates with LlamaIndex query engines and supports multiple correction strategies without requiring manual intervention.

Combines multi-agent orchestration (via CrewAI) with RAG to enable iterative document interaction where agents can refine queries, request clarifications, and progressively build context across multiple retrieval cycles. Implements agent-driven retrieval where specialized agents (researcher, analyzer, synthesizer) coordinate to decompose complex questions into sub-queries, retrieve relevant documents for each sub-query, and synthesize results. Uses LlamaIndex for document indexing and CrewAI for agent coordination, enabling complex reasoning patterns like hypothesis testing and evidence gathering.

Integrates speech recognition (via AssemblyAI or similar) with RAG to enable voice queries and voice-based document interaction while preserving audio context like speaker tone and emphasis. Converts speech to text with speaker diarization and confidence scores, then routes to RAG pipeline with audio metadata attached. Supports voice output via text-to-speech, enabling fully conversational document interaction. Implements streaming audio processing for real-time transcription and retrieval.

Implements a specialized multi-agent system (via CrewAI) for financial analysis where agents have access to domain-specific tools (financial data APIs, calculation engines, visualization tools) and coordinate to analyze financial documents, market data, and company information. Each agent has a specific role (analyst, researcher, report generator) with access to tools like stock price APIs, financial statement parsers, and calculation engines. Agents collaborate through task definitions and context sharing to produce comprehensive financial reports.

Implements an autonomous agent (via CrewAI) that can browse the web in real-time to retrieve current information, answer questions about recent events, and gather data from online sources. Uses Stagehand or similar browser automation to navigate websites, extract information, and synthesize findings. Agents can follow links, fill forms, and interact with dynamic content to gather information that isn't available in static documents. Integrates with RAG for combining web-retrieved information with local documents.

Provides reference implementations of Model Context Protocol (MCP) servers that standardize tool integration across different LLM providers and clients. Implements MCP server patterns for KitOps, SDV, and audio analysis tools, enabling any MCP-compatible client to access these tools through a standardized interface. Handles schema definition, request/response serialization, and error handling according to MCP specification. Supports both stdio and HTTP transport protocols for flexible deployment.

Provides a framework for comparing LLM models (GPT-4, Qwen3, open-source models) on specific tasks using Opik for experiment tracking and custom evaluation metrics. Implements evaluation pipelines that run the same prompts against multiple models, collect outputs, and score them using task-specific metrics (BLEU, ROUGE, custom domain metrics). Tracks experiments with full reproducibility including model versions, prompts, and hyperparameters. Integrates with OpenRouter for multi-model access.

Implements document understanding using multimodal vision models (Llama 3.2 Vision, Gemma-3) to extract text, tables, and structured data from images and PDFs. Processes documents through vision models that understand layout, tables, and formatting, then extracts structured data (JSON, CSV) from visual content. Supports batch processing of document collections and integrates with RAG for indexing extracted content. Handles complex layouts including multi-column text, tables, and mixed content.