Lantern

Agentic Business Intelligence Platform powered by LangGraph

Lantern is an advanced, multi-agent AI analytics platform designed to autonomously ingest, analyze, and visualize complex datasets. By leveraging a sophisticated LangGraph architecture with dynamic human-in-the-loop checkpoints, Lantern bridges the gap between autonomous code generation and executive-level business intelligence.

---

Demo

Important

Access the Live Platform

Dashboard

Analysis Workflow

Generated Report

Semantic Search

---

Key Features

• Multi-Agent LangGraph Workflow: A state-machine-driven orchestration of specialized agents handling distinct phases of the data lifecycle.
• Human-in-the-Loop Planning: The system proposes structured analysis plans that require human review and approval before execution, ensuring alignment and safety.
• Dynamic Python Code Generation: Autonomous generation of pandas and plotly code based on the approved analysis plan.
• Secure Sandbox Execution: Isolated execution of generated Python code to prevent side-effects and guarantee operational security.
• Interactive Plotly Visualizations: Dynamic, interactive charts embedded directly into the final executive reports.
• Executive-Level Business Insights: High-level interpretations of data metrics distilled for non-technical stakeholders.
• Recommendation Generation: Actionable, data-driven business recommendations produced by the Insight Agent.
• PDF & Markdown Reports: Beautifully formatted reports exportable for offline sharing and archiving.
• Semantic Memory using pgvector: Historical analyses are embedded and stored in a vector database for long-term intelligence retention.
• Historical Analysis Search: Natural language retrieval of past reports, enabling cross-dataset insights.
• Workflow Observability Dashboard: A React Flow interface providing real-time visualization of agent state transitions and graph execution.
• Multi-LLM Support: Built-in integrations for Gemini 2.5 Flash and OpenRouter (e.g., Nemotron 3), allowing flexible model selection.

---

Architecture

graph TD
    User([User]) --> |Uploads Dataset & Prompts| Frontend
    
    subgraph Frontend [Frontend Application]
        UI[Next.js + TailwindCSS]
        Obs[React Flow Observability]
    end
    
    Frontend --> |API Requests| Backend
    
    subgraph Backend [FastAPI Backend]
        API[FastAPI Routers]
        API --> LangGraph[LangGraph State Machine]
        
        subgraph LangGraph_Orchestration [LangGraph Orchestration]
            DU[Dataset Understanding Agent] --> DQ[Data Quality Agent]
            DQ --> AP[Analysis Planning Agent]
            AP --> |Interrupt| HIL{Human Approval}
            HIL --> |Resume| EX[Execution Agent]
            EX --> VZ[Visualization Agent]
            VZ --> IN[Insight Agent]
            IN --> RG[Reporting Agent]
        end
        
        LangGraph -.-> |State Persistence| CP[(Neon PostgreSQL Checkpoints)]
    end
    
    EX --> |Sandboxed Code| Sandbox[Python Sandbox]
    RG --> |Embed & Store| PGV[(pgvector Semantic Memory)]
    
    subgraph AI_Providers [AI Providers]
        Gemini[Gemini 2.5 Flash]
        OpenRouter[OpenRouter / Nemotron]
        Embed[text-embedding-004]
    end
    
    LangGraph_Orchestration <--> AI_Providers

Loading

---

Agent Workflow

flowchart LR
    A[Dataset Upload] --> B(Dataset Understanding)
    B --> C(Data Quality Review)
    C --> D(Analysis Planning)
    D --> E{Human Approval}
    E --> F(Analysis Execution)
    F --> G(Visualization)
    G --> H(Insight Generation)
    H --> I(Recommendation Generation)
    I --> J(Report Generation)
    J --> K[(Semantic Memory)]

Loading

---

Technology Stack

Category	Technologies
Frontend	Next.js, TypeScript, TailwindCSS, shadcn/ui, React Flow, Plotly
Backend	FastAPI, LangGraph, LangChain
Database	PostgreSQL, Neon, pgvector
AI	Gemini 2.5 Flash, OpenRouter, text-embedding-004
Infrastructure	Vercel, Render

---

Human-in-the-Loop Intelligence

Fully autonomous agents often hallucinate or execute misaligned objectives when faced with complex, open-ended datasets. Lantern solves this by implementing a Human-in-the-Loop (HITL) architecture:

1. The AI evaluates the dataset and generates a proposed analysis plan.
2. The LangGraph workflow intentionally interrupts and persists its state to PostgreSQL.
3. The user reviews, modifies, or directly approves the plan via the frontend.
4. Upon approval, the graph resumes execution.

This approach guarantees that the computationally expensive code generation and execution phases are strictly aligned with user intent, maximizing efficiency and safety.

---

Semantic Memory

Lantern acts as a long-term intelligence hub for your organization.

• Embedding: Final reports and insights are vectorized using Google's text-embedding models.
• Storage: Vectors and metadata are stored in a Neon PostgreSQL database utilizing the pgvector extension.
• Retrieval: Users can perform natural language semantic searches to instantly recall insights from past analyses, enabling historical context to inform future decisions.

graph LR
    Report[Final Report] --> Embed[Embedding Model]
    Embed --> |Vector| PGV[(pgvector Database)]
    Query[User Query] --> Embed2[Embedding Model]
    Embed2 --> |Similarity Search| PGV
    PGV --> Results[Historical Insights]

Loading

---

Observability

Understanding why an AI made a decision is just as important as the decision itself. Lantern features a dedicated observability dashboard built with React Flow.

This interface provides real-time, node-level visibility into the LangGraph state machine, allowing engineers to trace execution paths, inspect agent inputs/outputs, monitor state transitions, and seamlessly debug complex multi-agent interactions.

---

Example Analysis Lifecycle

1. Dataset Upload: A sales dataset is ingested and parsed.
2. Analysis Plan Generation: The AI identifies key metrics (e.g., MoM growth, regional performance).
3. Human Approval: The stakeholder reviews the plan and approves.
4. Execution: Python pandas code is generated and executed in a sandbox to compute the metrics.
5. Visualization: Interactive Plotly charts are generated for the metrics.
6. Insight Production: An LLM interprets the trends from the executed data.
7. Recommendation Production: Actionable business steps are proposed.
8. Executive Report: A beautifully formatted Markdown/PDF report is compiled.
9. Semantic Memory: The report is embedded and stored for future retrieval.

---

Engineering Highlights

• Orchestrated a complex, multi-agent state machine using LangGraph to handle deterministic and non-deterministic workflows.
• Implemented resilient, persistent graph checkpoints using PostgreSQL, enabling seamless interruptions and resumptions.
• Designed a scalable vector-based semantic retrieval system using pgvector for long-term conversational memory.
• Engineered a dynamic code generation and secure sandbox execution pipeline for real-time Python execution.
• Built interactive, responsive business intelligence dashboards integrating Plotly and React Flow.
• Developed a robust observability and tracing layer for deep agent visibility.

---

Future Roadmap

• Live workflow streaming
• Collaborative workspaces
• Multi-dataset analysis
• Scheduled reports
• Advanced agentic RAG
• Enterprise authentication

---

Why Lantern?

Lantern represents the cutting edge of applied AI engineering. By combining the autonomous capabilities of Agentic AI with the structured, deterministic requirements of Business Intelligence, Lantern delivers a platform that is both powerful and reliable.

The integration of Semantic Memory ensures that the system grows smarter over time, while Human-in-the-Loop checkpoints guarantee safety and alignment. Orchestrated by LangGraph, Lantern is not just a demo—it is a production-grade blueprint for the future of AI analytics.