Unscript One — AI-Native Enterprise Workspace

Problem

Organizations building AI internally repeatedly hit the same wall: chatbots that hallucinate, no visibility into whether answers are grounded, no control over tool access, and no way to measure quality over time.

Solution

Unscript One delivers the smallest complete version of the real enterprise AI problem: grounded answers via hybrid RAG, controlled tool access through MCP, measurable quality via automated Ragas evaluation, and full observability on one dashboard.

Timeline

RAG pipeline, LangGraph agents, MCP integration, guardrails, Ragas eval loop, Next.js workspace, status page.

Features

• Enterprise RAG with recursive chunking, hybrid retrieval, reranking, and citation-based responses.
• AI Research and Coding Agents powered by LangGraph with real GitHub MCP integration.
• Prompt injection and PII guardrails running pre-generation with visible pass/fail badges.
• Automated LLM evaluation using Ragas with faithfulness and answer relevance metrics.
• AI Observability Dashboard and public Service Health Monitoring at /status.
• Multi-session AI conversations with document ingestion and knowledge management.

Challenges

• Building a production-ready hybrid retrieval pipeline that balances BM25 recall with vector search precision.
• Designing LangGraph agent flows that expose tool call traces transparently in the UI.
• Running Ragas evaluation automatically on every deploy without blocking the release pipeline.
• Keeping prompt injection and PII guardrails fast enough to run pre-generation without adding noticeable latency.

Results and signals

• Faithfulness and answer relevance scores measured automatically on 20 hand-written test queries.
• Public /status page showing real uptime, request count, last eval timestamp, and current faithfulness score.
• Hybrid retrieval combining BM25 and vector search for higher recall than either alone.
• One real MCP tool call to GitHub — not a mock.

Future roadmap

• Add output-side guardrails: hallucination detection and sensitive data masking.
• Introduce a second agent and basic multi-agent handoff via LangGraph.
• Expand evaluation coverage with DeepEval and LangSmith integration.