Inspiration
I build Power BI dashboards myself in my day-to-day, and now that Power BI has become programmable (PBIP/TMDL), there’s a huge opening for AI to generate production-grade analytics—not just mockups. Plenty of “AI dashboard” tools exist, but almost none plug into Power BI, the platform non-technical business users already trust. SMONE CLI is about breaking that glass ceiling: bring agentic AI to the actual BI stack people use every day.
What it does
- Turns a CSV/Excel + a business prompt into a ready-to-open Power BI project (PBIP).
- Profiles data, proposes a semantic model (tables, measures, relationships), and generates a first-pass Report with usable visuals.
- Brand-ready output: supports adding your company logo and applying a custom color palette/theme directly in the generated report.
- Runs locally/offline (GPT-OSS 20B via Ollama) for privacy and portability.
- Outputs a clean PBIP/TMDL structure you can commit, diff, and iterate on with your team.
How we built it
- Multi-agent pipeline: Data Profiler → Schema/Intent Extractor → Model Builder (TMDL) → Report Planner → Report Writer (PBIR/JSON).
- Programmatic Power BI: emits PBIP/TMDL + PBIR assets (report.json, themes, versioning).
- Local LLM loop: GPT-OSS 20B (Ollama) for schema inference, measure suggestions, and visual planning with tight guardrails.
- Deterministic writers: everything compiles to files—no hidden state, no vendor lock-in.
Challenges we ran into
- Bridging “AI ideas” with Power BI realities (field well-formedness, visual bindings, measure signatures).
- Keeping the LLM creative but consistent (strict JSON schemas, validation passes, post-processing).
- Ensuring the pipeline stays modular (domain vs infra) so teams can swap LLMs or writers without rewiring everything.
Accomplishments that we're proud of
- A single command that produces a real PBIP you can open in Power BI—today.
- Offline-first by design: useful for regulated environments and confidential datasets.
- A clean architecture that separates orchestration, domain rules, and file I/O, making contributions and audits easy.
What we learned
- The moment BI becomes text-addressable and file-based, AI stops being a demo and becomes a tooling layer.
- Guardrails (validators, post-processors, mappings) matter as much as the model—quality = AI × rules.
What’s next for SMONE CLI
- Slide-deck generation from the report (auto-narratives, key visuals, and executive summaries).
- Benchmarks & evals for semantic models and report quality (repeatable tests, not vibes).
Built With
- gpt-oss-20b
- ollama
- python
Log in or sign up for Devpost to join the conversation.