Inspiration
Over 80% of enterprise AI automation projects fail before ever reaching production. The reason is rarely the underlying foundation model—it is the Executive Delusion Gap. Leadership automates the process they assume they have (clean CRM pipelines, automated handoffs, 24-hour turnaround), completely blind to the messy frontline reality on the ground (unmonitored WhatsApp groups, shadow Google Sheets, and manual webhook backfills).
Traditionally, uncovering this reality requires hiring management consultants for 8 to 12 weeks at $150,000+, producing static slide decks that are obsolete the week they are delivered. We built Mimic to replace that entire consulting cycle with autonomous AI discovery: deploying unscripted employee interview probes, mapping the actual shadow process DAG, flagging AI landmines before a single line of automation code is written, and outputting production-ready Model Context Protocol (MCP) agents in minutes.
What it does
Mimic is an autonomous enterprise operational discovery and alignment engine:
Dynamic Intake & Persona Generation: Ingests an executive mandate and dynamically provisions parallel stakeholder interview personas (Sales Rep, Customer Success, RevOps Lead, IT Ops). Parallel Workforce Probes & Live Employee Interview Terminal: Autonomous agents conduct unscripted, psychologically astute interviews with frontline staff to extract hidden workarounds, off-book tooling, and tribal work rules. Dual-DAG Synthesis & The Delusion Dashboard: Constructs a side-by-side comparison between the Presumed Executive Process (a clean linear line) and the Frontline Shadow Reality (a looping, high-friction graph). It calculates a mathematical Delusion Index (0–100%) and classifies workflows into AI Landmines (brittle processes that will crash if automated today) and AI Goldmines (immediate, high-ROI automation targets). Live Technical Feasibility Auditing: Integrates real-time web search to verify third-party API rate limits, schema constraints, and technical failure modes before approving an automation. 1-Click FastMCP Agent Blueprint Export: Generates production-ready FastMCP Python servers, system prompts, guardrails, and Claude Desktop configs targeting the validated golden workflow.
How we built it
LLM Inference & Intelligence: Powered by Nebius Token Factory utilizing NVIDIA Nemotron-3 120B (nvidia/nemotron-3-super-120b-a12b) for unscripted stakeholder probing, complex cross-transcript discrepancy reconciliation, dual-DAG node synthesis, and code generation. Live Search Grounding: Integrated Tavily Search API (AsyncTavilyClient) to conduct real-time feasibility audits against external developer docs and API rate limits. Backend Architecture: Built with FastAPI (Python 3.13) with asynchronous concurrency, Pydantic validation, and an embedded aiosqlite database (strata.db) storing sessions and synthesis artifacts. Agent Export Layer: Implemented the open Model Context Protocol (MCP) using FastMCP, compiling discovery findings into typed Python @mcp.tool() servers. Frontend & Visual Experience: Developed with React 19, Vite 8, and HeroUI v3, styled with Tailwind CSS v4 featuring a bespoke dark-gold luxury palette (#c5a059), glassmorphic panels, and animated Dual-DAG process nodes.
Challenges we ran into
Prompt Engineering for Unscripted Discovery: Early prompt iterations tended to follow generic survey questionnaires. We had to calibrate Nemotron-3 120B's system instructions to act like an aggressive, empathetic investigative auditor—detecting evasions, asking follow-ups about undocumented tooling (e.g. WhatsApp, private spreadsheets), and extracting exact tool names. Dual-DAG Structural Extraction: Converting unstructured, contradictory conversational interview transcripts from 4 different stakeholders into two strictly aligned directed acyclic graphs required careful JSON schema constraints and multi-pass reconciliation logic. Discrepancy Reconciliation: Reconciling the executive's belief that an API sync takes "24 hours" with frontline reality showing manual rework takes "48+ hours" required designing a robust friction node scoring algorithm to quantify the Delusion Index.
Accomplishments that we're proud of
The "Aha Moment" Dual-DAG Visualizer: In under 10 seconds, users immediately see why enterprise AI projects fail: the stark contrast between leadership's clean straight line and the messy, looping frontline reality with glowing red friction nodes and clickable quote evidence. True Unscripted Interview Experience: Frontline employees can open their private link and have a natural, adaptive conversation where the agent digs deeper into specific operational landmines. From Discovery to Code in Minutes: Closing the loop from problem discovery to code generation by producing a fully working FastMCP server and Claude Desktop config that users can immediately run. Production-Grade Aesthetics: Designing an interface that feels like high-end enterprise software rather than a prototype.
What we learned
The Delusion Gap is universal: Almost every enterprise workflow has an unmonitored shadow process. AI automations built without discovery are doomed to automate broken assumptions. Nemotron-3 120B on Nebius Token Factory excels at enterprise synthesis: The model demonstrated remarkable capability at digesting dense, multi-stakeholder interviews, extracting subtle contradictions, and synthesizing strict DAG topologies and typed Python code with high fidelity and ultra-low latency. MCP is the ideal deployment bridge: Generating standard MCP servers allows operational discovery tools to interface directly with the next generation of agentic IDEs and desktop assistants like Claude.
What's next for Mimic
Live Voice Discovery Interviews: Adding real-time voice-to-voice interview probes so frontline workers can complete their discovery check-in over a quick 2-minute phone call. Automated Workflow Shadowing: Integrating with browser extensions and desktop audit logs (Slack, email, ERP) to corroborate interview claims against actual telemetry. Enterprise MCP Registry Integration: Direct 1-click deployment to enterprise MCP gateways and cloud agent runtimes.
