Premiering at DATA+AI Summit 2026

Marvin Nahmias  &  Javier Hauss  · Alpura · Mexico City

Alpura Alpura Databricks Databricks

AI Genie Factory

Build enterprise Databricks apps in days. Not months.

|
Get Started View on GitHub
0 Apps Shipped
0 Agent Skills
0 Chars of Constraints
0 Business Units
Architecture

The Factory Methodology

Every layer constrains the one above it. AI generates code inside those constraints — it never defines the architecture.

Platform (Lakehouse) Foundation
Governance (Unity Catalog) Layer 2
Semantic Layer (Gold Tables) Layer 3
Factory (AGENTS.md + Skills) This Repo
Genie Code Agent Mode Layer 5
Apps / Pipelines / Notebooks Output
The core insight

"AI doesn't generate architecture — it generates code. Architecture has to be defined first, then enforced as constraints."

— Marvin Nahmias, DAIS 2026
Results at Alpura
Apps shipped 90+
Business units Finance, Sales, Ops, Marketing
Typical build time Days, not months
Constraints

AGENTS.md vs Skills

Two complementary layers. One keeps every generation on-rails. The other loads domain expertise only when needed.

AGENTS.md
Always-on guardrails
When loaded Every single Genie Code session — automatic
Purpose Non-negotiable platform rules, stack, error contracts, logging
Size target < 6,000 chars — never bloated
Overridable Never. Platform law.
Skills
On-demand domain knowledge
When loaded Only when relevant — auto (Agent mode) or @mention
Purpose Charts, pipelines, testing, UX/UI design system
Size target As large as needed — depth over brevity
Overridable Per-app, via APP.md constraints section
Skills

Six Specialist Agents

Each skill is a focused context file deployed to your Databricks workspace. @mention any skill in Genie Code Agent Mode to load its expertise.

@data-access
Data Access
sql-connector + Config() credentials provider for Databricks Apps. Single-connection batching for performance. Decimal→float safety. Unity Catalog three-part names, DataAccessError contracts.
sql-connector Unity Catalog data.py
@ui-ux-patterns
UX/UI Patterns
Full design system — color tokens, shadow/elevation, Inter typography, KPI cards, Plotly dark theme, chart functions, Dash & Streamlit shells.
Design Tokens Plotly Dark Dash / Streamlit
@databricks-app
Databricks App
Full app architecture — data/logic/ui/app layers, app.yaml, service principal auth, OAuth token passthrough, deployment commands, debug checklist.
app.yaml OAuth Layer Architecture
@databricks-dashboard
AI/BI Dashboard
Native Lakeview dashboards — dataset SQL, parameterized filters, counter/chart/table tiles, layout templates, refresh scheduling, widget config.
Lakeview Parameterized SQL AI/BI
@dlt-pipeline
DLT Pipeline
Bronze/Silver/Gold medallion, Auto Loader, CDC with APPLY CHANGES, SCD Type 2, schema evolution, streaming vs batch, pipeline settings.
Bronze/Silver/Gold CDC / SCD2 Auto Loader
@testing-scaffold
Testing Scaffold
pytest patterns, mocked dbsql.connect() calls, pandas logic tests, DataAccessError assertion helpers. Tests business logic without hitting Unity Catalog.
pytest Mocked dbsql logic.py
Workflow

How It Works

Four steps from repo to running Databricks app. Most apps go from spec to shipped in a single Genie Code session.

01
Define Architecture
Edit GLOBAL_RULES.md, STACK.md, and modules. Run build_agents.py to compile AGENTS.md.
python build_agents.py
02
Deploy Skills
Push instructions and all 6 skills to your Databricks workspace with one command.
./deploy.sh
03
Fill Out APP.md
Copy APP_TEMPLATE.md, name your tables, define KPIs, describe the charts. One page per app.
cp templates/APP_TEMPLATE.md apps/my_app/APP.md
04
Prompt Genie Code
Open Agent Mode, paste your PROMPT_TEMPLATE + APP.md, @mention the skills you need.
@databricks-app-design build per APP.md
Setup

Deploy in Under 5 Minutes

Requires Databricks CLI v0.200+ and a configured profile. Everything else is a single shell script.

1
Install Databricks CLI
Install CLI v0.200+ and authenticate with databricks auth login.
2
Clone the Repo
Clone mexmarv/ai-genie-factory from GitHub and navigate to the root.
3
Run deploy.sh
Uploads AGENTS.md as your instructions file and deploys all 6 skills to your workspace.
4
Open Genie Code
Click the sparkle icon in Databricks. Switch to Agent Mode. Your skills are live.
terminal 
# 1. Authenticate (one-time)
databricks auth login

# 2. Clone
git clone https://github.com/mexmarv/ai-genie-factory
cd ai-genie-factory

# 3a. Deploy to your personal .assistant folder
./deploy.sh

# 3b. Deploy workspace-wide (admin required)
./deploy.sh --workspace

# 3c. Use a specific CLI profile
./deploy.sh --profile my-profile

# 4. Rebuild AGENTS.md after editing modules
python build_agents.py

deploy.sh uses --format RAW to upload .md files — import-dir only handles notebooks.

Examples

Ready-to-Run Apps

Two APP.md specs included. Start with NYC Taxi Explorer to verify the factory works before building on your own data.

Zero Permissions
NYC Taxi Explorer
samples.nyctaxi.trips
5-component dark dashboard with KPI cards, daily volume line chart, trips by hour, payment type treemap, and distance vs fare scatter. Works in every Unity Catalog workspace out of the box.
Requires System Tables
DBU Spend Monitor
system.billing.usage
Track Databricks Unit consumption by day and cluster. KPI card for grand total, line chart for daily trend, horizontal bar for top 10 clusters. Requires system tables enabled by an admin.