Autocrisp: Agentic Data Mining Platform

Technical showcase & architecture specification
This repository outlines the system architecture and technical design of Autocrisp. Because of intellectual property agreements and non-disclosure requirements, the production source code remains in private institutional repositories.

---

Overview

Autocrisp is a low-code orchestration platform that guides users through data science workflows without writing code. It helps researchers and domain specialists frame questions, clean datasets, train predictive models, and interpret results through conversational AI agents.

The system follows the CRISP-DM (Cross-Industry Standard Process for Data Mining) lifecycle. Instead of handing the entire pipeline to an unconstrained code-generation model, Autocrisp assigns dedicated agents to discrete phases. It requires human sign-off during early problem scoping and data cleaning, keeping flawed assumptions from cascading into modeling.

---

Institutional backing & publications

Autocrisp was developed by Jack van der Vall (AI Engineering Intern) in collaboration with the Erasmus Centre for Data Analytics (ECDA) at Erasmus University Rotterdam, and the Applied Data Science & AI program at Hogeschool Rotterdam.

The platform was developed and tested in the EDC Data Sandbox, the data and compute environment managed by the Erasmus Data Collaboratory (House of AI).

• Official portfolio case study: ECDA: Autocrisp: AI Agents that Automate CRISP-DM
• Institution: Erasmus Centre for Data Analytics (ECDA) / Erasmus Data Collaboratory | House of AI
• Academic partner: Applied Data Science & AI, Hogeschool Rotterdam
• Developer / AI intern: Jack van der Vall (LinkedIn)

---

Technical architecture

Autocrisp separates deterministic workflow automation from autonomous agent reasoning:

flowchart LR
    subgraph Frontend["Frontend Layer"]
        UI["Streamlit UI<br/>(Conversational & Visual)"]
    end

    subgraph Orchestration["Deterministic Orchestration Layer"]
        N8N["N8N Workflows<br/>(Webhooks & API Connectors)"]
        OrchAgent["Orchestration Coordinator"]
        N8N --> OrchAgent
    end

    subgraph Business["Phase 1: Business Understanding"]
        QueryAgent["Query Refiner Agent"]
        WebAgent["Web Research Agent"]
        DataDiscAgent["Dataset Discovery Agent"]
        CompilerAgent["Research Compiler Agent"]
        QueryAgent --> WebAgent --> DataDiscAgent --> CompilerAgent
    end

    subgraph DataOps["Phases 2-5: Data Ops & Modeling (AG2 / AutoGen)"]
        ExplorerAgent["Data Explorer Agent<br/>(Automated EDA)"]
        ProcessorAgent["Data Processor Agent<br/>(Sanitization & Features)"]
        TrainerAgent["Model Trainer Agent<br/>(scikit-learn / PyTorch)"]
        CodeExec["Sandboxed Code Executor<br/>(UserProxy Container)"]
        ExplorerAgent --> ProcessorAgent --> TrainerAgent --> CodeExec
    end

    subgraph Backend["State & Security (Supabase)"]
        Auth["User Auth (RLS)"]
        Credentials["Encrypted Credentials"]
        Memory["Agent Memory / Research Logs"]
        Storage["Datasets & Artifacts"]
    end

    subgraph External["Models & External APIs"]
        LLMs["Multi-LLM Providers<br/>(OpenAI / Gemini / DeepSeek)"]
        SearchAPI["Web Search API<br/>(Perplexity / Sonar)"]
    end

    UI -->|"Natural Language Query"| N8N
    UI -.->|"Session Auth"| Auth
    OrchAgent -->|"Delegation"| QueryAgent
    OrchAgent -->|"Structured Pipeline Task"| ExplorerAgent
    CompilerAgent -.->|"Persist Findings"| Memory
    ExplorerAgent -.->|"EDA Insights & Charts"| UI
    DataOps -.->|"Model Inference & Synthesis"| LLMs
    Business -.->|"Web Search"| SearchAPI
    CodeExec -.->|"Artifacts & Metrics"| Storage

Loading

Click to view high-resolution architecture diagram

---

The CRISP-DM agent funnel

Autocrisp maps the six CRISP-DM phases to dedicated agent roles:

1. Business understanding (N8N research workflows)

The process starts with research agents that convert plain-text goals into clear hypotheses. Before touching raw data, the agents use web search (via Perplexity) to gather domain context, explore public benchmarks, and discover relevant external datasets.

2. Data understanding (Data Explorer agent)

Once a dataset is uploaded, the explorer agent profiles distributions, checks column correlations, flags missing values, and produces interactive Plotly and Seaborn charts inside Streamlit.

3. Data preparation (Data Processor agent)

The processor agent cleans dirty records, handles imputations, resolves collinearity, and engineers new features informed by the research phase. Every transformation produces a verifiable Pandas script before execution.

4. Modeling (Model Trainer agent)

The trainer agent configures classification, regression, or clustering pipelines using scikit-learn, XGBoost, or PyTorch. It tests model candidates, runs hyperparameter sweeps, and logs cross-validation results.

5. Evaluation (Evaluation agent)

Rather than just returning raw scores, this agent explains model performance and confusion matrices in plain language so users can evaluate results before deciding to iterate or deploy.

6. Deployment & export

Users can export clean, containerized Python scripts, trained model files (.pkl or .onnx), and step-by-step audit logs to hand off to engineering teams.

---

Core technical stack

Component	Technology	Role
Agent framework	AG2 (AutoGen)	Conversable loops, automated code generation, and iterative syntax repair.
Orchestration	N8N	Deterministic workflows for webhooks, web search integration, and API queries.
Memory & storage	Supabase (PostgreSQL)	Conversation state, research logs, Row-Level Security (RLS), and encrypted credentials.
User interface	Streamlit	Interactive UI for agent dialogues, data visualizations, and approval gates.
Code execution	Docker sandbox containers	Isolated runtime for running generated Pandas, scikit-learn, and PyTorch scripts without host exposure.
LLM routing	Multi-provider dispatch	Routing between GPT-4o, Gemini Pro, and DeepSeek based on task complexity.

---

Engineering and security controls

• Sandboxed code execution: All code proposed during agent loops runs in isolated Docker containers. The host machine is isolated from generated scripts, blocking unintended filesystem or network access.
• Human-in-the-loop review: Agents cannot run data transformations or train models without explicit user confirmation. Users review proposed code and data edits before execution begins.
• Row-Level Security (RLS) & data isolation: User accounts, datasets, and project data are secured in Supabase with PostgreSQL Row-Level Security, enforcing strict access control.
• Stateful memory & session persistence: Agents retain context across the project lifecycle via encrypted storage in Supabase, keeping previous research findings, data cleaning steps, and generated scripts accessible across sessions.

---

Contact & project links

For research inquiries, demonstrations, or collaboration:

• Developer: Jack van der Vall (LinkedIn)
• Portfolio case study: Erasmus Centre for Data Analytics