Technical Content Writer Agentic Platform

Deep Blog Agent (Technical Content Writer Agent) is a Streamlit + LangGraph based agentic AI system designed to generate high-quality technical blog posts from a short prompt or topic. It combines structured planning, web research, long-form content synthesis, and optional image generation into a single reproducible workflow.

Overview

This project solves a common problem in technical writing:

How to quickly generate a structured, research-backed draft while maintaining transparency and reproducibility.

The system provides:

• A Streamlit workspace with live workflow visibility
• A LangGraph-based multi-stage pipeline
• Optional web research using Tavily
• Optional image generation for diagrams
• Persistent artifact storage for reproducibility

System Workflow

Each run follows a structured pipeline:

1. Accept a topic or prompt
2. Decide whether research is needed
3. Retrieve and process evidence
4. Generate structured outline
5. Draft sections in parallel
6. Assemble final markdown
7. Optionally generate and embed images
8. Save outputs as reusable artifacts

Architecture (Agentic Pipeline)

The system is implemented using LangGraph, enabling modular and observable workflows.

Router → Research → Orchestrator → Workers → Reducer

• Router: decides research vs closed-book mode
• Research Node: retrieves external knowledge (Tavily)
• Orchestrator: builds structured plan (Pydantic schema)
• Workers: generate sections in parallel
• Reducer: combines outputs into final blog

Key Features

Agentic Blog Generation

• Multi-stage structured reasoning pipeline
• Evidence-grounded content generation
• Modular section-wise drafting
• Deterministic output formats

Research Integration

• Tavily-powered web search
• Citation-aware generation
• Up-to-date knowledge incorporation

UI & Observability

• Live workflow execution tracking
• Run history and replay
• Detailed run inspection (logs, sources, outputs)
• Transparent agent reasoning

Artifact System

Each run generates reusable outputs:

outputs/<timestamp>_<slug>/
  ├── blog.md
  ├── run.json
  └── images/

• Markdown export
• Metadata tracking
• Image bundling
• Reproducibility support

FinOps & Cost Awareness

• Token usage tracking
• Cost estimation per run
• Configurable pricing assumptions

Observability & Monitoring

Streamlit Interface

The UI is structured into multiple functional pages:

• Home → Prompt input + live workflow
• Prompt Examples → reusable prompts
• Run History → previous sessions
• Run Detail → full trace + outputs
• Settings → API keys + model config
• FinOps → cost tracking

Tech Stack

• Backend / Agents: LangGraph, LangChain
• LLMs: OpenAI (text), Google GenAI (images)
• Research: Tavily API
• Frontend: Streamlit
• Data Models: Pydantic
• Environment: Python 3.13, uv

CLI Usage

Supports both UI and CLI workflows:

deep-blog-agent “Write a technical blog on RAG evaluation in production”

Optional flags:

• –no-research
• –no-images
• –print-markdown

Implementation Highlights

• LangGraph-based modular orchestration
• Parallel section generation
• Schema-driven planning with Pydantic
• Separation of concerns (providers, core, agents, UI)
• Reproducible artifact pipeline
• Session-based API key handling in UI

What I Learned

• Designing agentic workflows with structured reasoning
• Building observable AI systems (LangSmith + logs + UI tracing)
• Integrating retrieval + generation + planning pipelines
• Managing cost-aware GenAI systems (FinOps layer)
• Creating production-ready ML/LLM applications with UI + CLI

Links

• GitHub
• Live Demo
• Demo Video