Atlas Oracle

Inspiration 🌍✨

Gemini 3 Flash made one thing undeniable: fast, grounded reasoning is finally usable at human speed. That pushed us to rethink how location intelligence should work. Instead of dashboards, layers, and gut feel, what if a location could explain itself end to end using evidence, constraints, and economics. Atlas Oracle was born from treating geography as an input to a reasoning system, not a visualization problem.

What it does 📍🧠

Atlas Oracle turns a single coordinate into a structured feasibility brief. It identifies concrete business opportunities, invents plausible project concepts, estimates startup costs, and surfaces risks.

• The Brain: Gemini 3 Flash reasons over real-time evidence.

• The Voice: An ElevenLabs agent allows you to "interrogate" the analysis, defending its logic and exploring risks through natural conversation.

How we built it 🏗️⚙️

Atlas Oracle is a browser-first reasoning pipeline disguised as a UI built in Google AI Studio.

The frontend orchestrates multiple public geospatial data sources in parallel. Overpass API scans the surrounding area for economic signals like amenities, land use, transport, and commercial density. Nominatim adds administrative and zoning context. Open-Meteo supplies elevation and environmental constraints.

All raw responses are aggressively normalized into a compact, token-efficient context object. Noise is stripped. Signals are preserved. This normalized context is then passed into Gemini 3 Flash with strict output constraints and search grounding enabled.

Gemini’s role is not to hallucinate ideas, but to reason over evidence. It infers economic tier, demand gaps, and feasibility, then returns a tightly structured JSON schema containing opportunities, risks, assumptions, and cost logic.

That same structured output becomes the single source of truth across the app. It renders visual cards, drives PDF generation, and is injected directly as dynamic variables into an ElevenLabs conversational session. The voice agent does not re-analyze anything. It explains, defends, and explores the already-grounded reasoning.

No backend orchestration. No vector databases. No async jobs. Just a deterministic, end-to-end reasoning loop running entirely in the browser.

Architecture 🏛️

Challenges we ran into 🚧

Public geospatial APIs are noisy, inconsistent, and rate-limited. We had to design resilient fallback strategies and aggressive data pruning. Gemini needed enough freedom to reason, but enough structure to stay grounded. Voice required careful scoping so it felt like an analyst, not a narrator. And everything had to work without a server safety net.

Accomplishments we’re proud of 🏆

We built a fully client-side location intelligence pipeline. We used Gemini 3 Flash for fast, explainable reasoning instead of slow multi-stage workflows. We treated structured JSON as the backbone of the entire product. And we integrated ElevenLabs as a reasoning interface, not a gimmick.

What we learned 📚

Reasoning quality improves dramatically when inputs are normalized, not raw. Smaller, faster models outperform larger ones when constraints are explicit. Voice becomes powerful when it is downstream of analysis, not upstream. And users trust AI more when the data path is visible and coherent.

What’s next for Atlas Oracle 🚀

• Multi-Site Comparison: Comparative reasoning for regional planning.

• Hyper-Local Economics: Integrating local tax and labor cost APIs.

• Agentic Mobility: An ElevenLabs + WhatsApp integration for "on-the-ground" site analysis via voice.

Atlas Oracle proves that when Gemini 3 Flash reasoning is fed clean geospatial signals and paired with ElevenLabs conversational intelligence, locations stop being abstract. They become computable.

Built With

• elevenlabs
• gemini
• open-meteo
• openstreetmap
• overpass-openstreetmap
• react