TaskFlow AI

An AI Agent workflow automation platform MVP with real LangGraph-controlled execution, executable tool calls, human approval pause/resume, SSE live trace, usage logs, credits, replay, and an AgentOps dashboard.

Business value

Shows a higher-value agent engineering skill set beyond chatbots and RAG: stateful workflow execution, tool routing, approval gates, reliability limits, audit logs, and an honest demo-local fallback.

Key features

Guest demo loginReal LangGraph runtimeStateful agent runsMCP-style internal tool gatewayExecutable demo toolsHuman approval pause/resumeSSE live run trace

Next.js App RouterTypeScriptFastAPIPythonPostgreSQLSQLAlchemyAlembicLangGraph

live demo github

overview

TaskFlow AI is Project 2 in the portfolio: a flagship-level MVP focused on AI Agent workflow automation instead of another support dashboard or document RAG app.

problem

Many agent demos show a nice UI but hide the runtime. They do not prove stateful execution, tool calls, approval checkpoints, retries, cost controls, or run-level observability.

solution

Refactor the runtime so LangGraph controls the workflow path through load, permission validation, initialization, planner, tool router, tool execution, approval gate, agent execution, reviewer, finalize, and error handling nodes.

key features

Guest demo login / Real LangGraph runtime / Stateful agent runs / MCP-style internal tool gateway / Executable demo tools / Human approval pause/resume / SSE live run trace / Retry and error handling / Max step and cost limits / Usage logs and credits / Run replay endpoint / AgentOps dashboard / Technical review page

tech stack

Next.js App Router / TypeScript / FastAPI / Python / PostgreSQL / SQLAlchemy / Alembic / LangGraph / Pydantic / React Flow / Recharts / Tailwind CSS / Vercel / Render

architecture

Next.js provides the product UI and live run pages. FastAPI exposes workflow, run, approval, replay, usage, and gateway routes. SQLAlchemy persists workspaces, runs, run steps, tool calls, approvals, usage logs, credits, and replayable outputs.

AI-assisted workflow

The guest demo uses demo-local model behavior and seeded demo search so reviewers can run it without private API keys. External write tools are clearly simulated in demo mode, while internal tool execution and approval flow are real.

challenges and what I learned

The main learning was making the agent runtime reviewable: graph nodes mutate state, persist transitions, pause for approval from inside the graph, resume the same run, and keep usage and cost records honest.

screenshots

Public live demo is available at https://taskflow-ai-seven-eosin.vercel.app with guest login, workflow execution, approval pause/resume, and replay.