PROJECT OVERVIEW We are building a Skills module for an existing AI chatbot SaaS platform. A "Skill" is an automated workflow that the chatbot can trigger during a conversation — for example, looking up an order status via API, sending a notification email, or collecting and submitting form data. We need a developer to build this module as a standalone, isolated deliverable that we will integrate into our existing platform ourselves. CORE REQUIREMENTS 1. Skill Execution Engine (Backend) Build a DAG-based (directed acyclic graph) workflow engine that executes multi-step skills: Skill structure: A skill is a graph of actions (nodes) connected by edges (transitions). Each action has a type, configuration, and optional conditions for branching. Action types to support: - API Call — Make HTTP requests (GET/POST/PUT/DELETE) with configurable URL, headers, body, query params. Support variable interpolation from conversation context or previous action outputs. - AI Generate — Call an LLM (OpenAI API) with a prompt template to generate/transform text. The prompt may reference variables. - Data Transform — Apply JSONPath expressions, map/filter data, restructure JSON payloads between actions. - Knowledge Lookup — Query a vector database (Pinecone) with a search query to retrieve relevant documents. - Send Email — Send an email via SMTP or API (e.g., SendGrid) with template variables. - Webhook / Notify — Send a webhook POST to a configurable URL with a JSON payload. - Human Escalation — Flag the conversation for human handoff, pause skill execution. - UI Prompt — Present a message or choices to the user and wait for their response before continuing. Execution flow: - Topological sort of the DAG to determine execution order. - Support conditional branching: edges can have conditions (e.g., "if status == 'found', go to action B; otherwise go to action C"). - Actions can reference outputs of previous actions using a variable syntax like {{action_id.output.field}}. - Skill execution should be resumable (for actions that wait for user input). - Store execution state per session (which actions ran, their outputs, current position). Credential vault: Skills that call external APIs need stored credentials. Implement encrypted credential storage (AES-256-GCM) with: - CRUD operations for credentials (API keys, OAuth tokens). - OAuth2 authorization code flow support (initiate, callback, token refresh). - Credentials are scoped per bot/organization. 2. Skill Intent Detection Before a skill executes, the system needs to detect which skill to trigger based on the user's message: - Each skill has one or more intent phrases (example utterances). - Use embedding-based matching: embed intent phrases with OpenAI's text-embedding-3-small, store in a vector index. - At runtime, embed the user's message and find the closest matching skill above a configurable confidence threshold. - Return the matched skill (or none if below threshold). 3. AI-Powered Skill Builder (Backend) Build an AI agent pipeline that can generate complete skill configurations from a natural language description: User provides a prompt like: "When a customer asks about their order status, call our Shopify API to look up the order by email and return the status." The system should: 1. Analyze the prompt to determine required actions, API endpoints, data flow. 2. Generate a complete skill structure: actions, edges, configurations. 3. Auto-generate intent phrases for triggering. 4. Return the full skill JSON ready to save. Use a multi-agent approach: - An orchestrator LLM decomposes the request and plans the skill structure. - A builder LLM generates the detailed action configurations. Support iterative refinement: user can say "add error handling" or "also send a confirmation email" and the system modifies the existing skill. 4. MCP (Model Context Protocol) Integration Support connecting to external MCP servers that expose tools the chatbot can use: - Implement an MCP client that communicates via JSON-RPC 2.0 over Streamable HTTP and SSE (Server-Sent Events) transports. - Discover available tools from an MCP server (tools/list). - Execute MCP tools (tools/call) as skill actions. - Store MCP server configurations (URL, auth headers) per bot. 5. Visual Skill Builder (Frontend) Build a React-based visual editor for creating and editing skills: - Flow canvas: Use React Flow to render the skill graph. Users drag-and-drop action nodes onto a canvas and connect them with edges. - Node types: Each action type gets a distinct visual node with an icon, title, and status indicator. - Configuration panels: Clicking a node opens a side panel with a form to configure that action type (URL fields, header key-value editors, prompt text areas, JSONPath inputs, etc.). - Edge conditions: Users can click an edge to set branching conditions (simple condition builder with field/operator/value). - Skill settings panel: Name, description, intent phrases (add/remove), enable/disable toggle, and configuration for trigger conditions. - Test panel: A panel to simulate skill execution — user provides a test message, system runs the skill and shows step-by-step results (which actions ran, their inputs/outputs, timing). - Skill templates browser: A gallery of pre-built skill templates users can browse, preview, and instantiate. - AI generation UI: A dialog/wizard where the user types a natural language description and the system generates the skill, then loads it into the visual editor for review/tweaking. Frontend tech stack: React (Create React App), plain JavaScript (no TypeScript), plain CSS (no CSS frameworks), React Flow for the graph editor. 6. Skill Templates Provide a template system: - Templates are pre-configured skill blueprints (e.g., "Order Status Lookup", "Appointment Booking", "FAQ Escalation"). - Store templates with: name, description, category, required credential types, the full skill graph structure. - Users can browse templates by category, preview the flow, and create a skill from a template (which copies the structure and lets them customize). TECH STACK REQUIREMENTS - Backend language: Python 3.10+ - Backend framework: FastAPI - Database: PostgreSQL (via asyncpg, raw SQL — no ORM) - AI/LLM: OpenAI API (GPT-4.1 or later) - Embeddings: OpenAI text-embedding-3-small - Vector DB: Pinecone - Encryption: AES-256-GCM (Python cryptography library) - Frontend framework: React (Create React App) - Graph editor: React Flow - Styling: Plain CSS (no Tailwind, no CSS-in-JS) - Language: JavaScript (no TypeScript) - Primary accent color: #5616ea DELIVERABLES 1. Backend module — A self-contained Python package/directory with: - All route handlers (FastAPI routers) - Database models and migration SQL scripts (CREATE TABLE statements) - Execution engine - AI generation pipeline - MCP client - Intent detection index - Credential vault logic - Full API documentation (OpenAPI/Swagger or markdown) 2. Frontend module — A self-contained React component directory with: - Visual skill builder (React Flow-based) - All configuration panels - Test panel - Template browser - AI generation wizard - All required CSS files 3. Integration specification — A document describing: - All API endpoints (request/response schemas) - Database schema (tables, columns, types, foreign keys) - Environment variables / configuration needed - How the frontend components expect to communicate with the backend (API base URL, auth headers, etc.) 4. Docker demo — A docker-compose setup that runs the backend + a simple frontend demo page, so we can test the module in isolation before integrating. WHAT WE PROVIDE - Integration specification: We'll provide a document describing how skills plug into our existing chatbot flow — specifically: the data format of a chat message, how to call back into our system when a skill needs to send a response, and the authentication mechanism your API routes should expect. - Test API keys: OpenAI API key and Pinecone credentials for development/testing. - Design reference: Mockups or wireframes for the visual builder UI (if needed). WHAT WE DO NOT PROVIDE - Access to our existing codebase. This module must be built as a standalone, isolated deliverable. - Access to our production infrastructure. You'll develop and test locally. APPLICATION REQUIREMENTS In your proposal, please include: 1. Your experience with workflow/automation engines (DAG execution, state machines). 2. Your experience with React Flow or similar graph editor libraries. 3. Your experience with OpenAI API and building AI agent pipelines. 4. A rough architecture outline for how you'd structure the execution engine. 5. Estimated timeline with milestones. 6. Portfolio links or examples of similar work. Budget: Open to proposals. Please provide a fixed-price quote based on the scope above. Timeline: Flexible.