Master PM tech stack: AI assistant on FastAPI

A Step-by-Step Plan for PMs to Master Tech Stack by Building an AI Assistant

PMs without a technical foundation risk losing credibility with developers. Build a real SaaS project—an AI assistant based on a task tracker. It answers queries like "How's the team doing?" or "What's the developer's workload?" An alternative for beginners: an Open Source GitHub analyst that handles questions like "Critical bugs in React?"

The project covers client-server architecture, REST, unit tests, containerization, and logging. It requires live data and scaling to two frontend clients.

Tech Stack:

• Backend: Python, FastAPI, PostgreSQL, uv
• Frontend: React, TypeScript, Vite, Shadcn, Tailwind

Setting Up the Environment and Tools

Install Python, npm, pip (replace with uv), Git. Add an IDE (Antigravity), Docker Desktop, DBeaver. For AI coding, use Google AI Pro: frontier models (Gemini-3.1-pro high), IDE Antigravity, CLI Gemini.

MCP (Model Context Protocol) — Basic Set:

• serena
• sequential-thinking
• context7
• filesystem

MCP Config (mcp_config.json):

{
  "mcpServers": {
    "filesystem": {
      "$typeName": "exa.cascade_plugins_pb.CascadePluginCommandTemplate",
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "C:\\Users\\User\\a_projects",
        "C:\\Users\\User\\.ssh"
      ],
      "env": {}
    },
    "sequential-thinking": {
      "$typeName": "exa.cascade_plugins_pb.CascadePluginCommandTemplate",
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-sequential-thinking"
      ],
      "env": {}
    },
    "context7": {
      "$typeName": "exa.cascade_plugins_pb.CascadePluginCommandTemplate",
      "command": "npx",
      "args": [
        "-y",
        "@upstash/context7-mcp"
      ],
      "env": {}
    },
    "serena": {
      "$typeName": "exa.cascade_plugins_pb.CascadePluginCommandTemplate",
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/oraios/serena",
        "serena",
        "start-mcp-server"
      ],
      "env": {}
    }
  }
}

Install the skills.sh registry: npm install -g @skills/cli. Add skills as you progress through the stages.

Rules for Working with AI in Coding

Break down tasks: one chat per task. If it's not resolved, start a new chat with clarification. Write in English without roles or goals. For difficulties, apply the "duck" method:

• Discuss in chat, create context.md.
• In a new chat—vision.md for architecture.
• tasks.md with tasks.
• Execute tasks.

Stage 1: Telegram Bot with Jira via MCP

Create a bot on FastAPI: access Jira MCP, respond in Telegram. Tokens in .env. AI handles Atlassian MCP calls.

• Result: Local bot, basic chat.
• Stack: Python, FastAPI, Gemini API, Atlassian MCP.

Server: @aashari/mcp-server-atlassian-jira.

Stage 2: Direct REST Requests to Jira API

Abandon MCP. Git init, main branch. Scripts: users, statuses, project tasks (30 days).

Skill: fastapi-expert.

• Result: Control via HTTP, .venv, .gitignore, .env.example.
• Stack: Jira REST API, requests, Git.
• Practices: VCS, Tool Calling, API.

Stage 3: PostgreSQL with FTS

Schema: tasks, changelog. Sync button (overwrites old data). AI uses top-5 via FTS (tsvector/tsquery, BM25).

• Result: Local storage, fast search.
• Stack: PostgreSQL FTS, raw SQL.
• Practices: DB Design, SQL Migrations, IR.

Install tsvector, DBeaver.

Stage 4: Hybrid RAG with Vectors

Tables: assignee, status, type (int ID). Manual migrations. Vectorize tasks in Qdrant (HNSW). Search: top-5 tasks → JOIN changelog by task_id.

Indexes: assignee_id, status_id, type_id, HNSW, task_id. EXPLAIN ANALYZE for optimization.

Skill: postgresql-optimization.

• Result: Semantic search with history, performance on thousands of records.
• Stack: Qdrant, PostgreSQL (HNSW, tsvector).
• Practices: RAG, Hybrid Search, Vector DB, Query Opt.

Key Takeaways

• The project is built step-by-step: from MCP to direct APIs, SQL, RAG.
• Focus on practices: VCS, indexing, JOIN, query optimization.
• AI accelerates, but the PM controls architecture and stack.
• Extensions: two frontends, server deployment, AI integration.

— Editorial Team