Altitude

Inspiration

Every hard questions lives at several altitudes at once. A production bug is a symptom, there might be three competing explanations and there might be one log line that explains everything.

A market call is macro view, but it might rely on earnings on a Thursday or some random tweet.

The understanding doesn't come from holding any one of these ideas, but instead moving between them without losing your place.

What it does

Altitude is a thinking environment: big framings on top, the claims that would make them true beneath, real evidence at the ground. When evidence lands, it climbs, flips the claim, re-scores the thesis.

How we built it

The heart is one pure Python file: a deterministic engine that holds the field. Everything else is a thin adapter around it.

Around the engine:

• SQLite for workspaces, with an outbox to Raven, our memory graph (Postgres + pgvector).
• One port serving everything: the browser field, a JSON API, and MCP — so humans, voice agents, and coding agents all operate the same field through the same boundary.
• A single-file browser UI, no framework. A CLI that talks HTTP like everyone else — it never imports the engine, so our automation tests the real surface.
• LLM inference through one OpenAI-compatible client (Fireworks by default, anything via env), plus a deterministic offline stub for tests and demos — never as a silent fallback.
• Docker end to end: the app, the memory stack, and a disposable fleet of N instances we can spin up, stress, and throw away.

Challenges we ran into

• Evals were harder than the build.
• Iteration speed on UX design.
• Finding the right level of abstraction — for a product about levels of abstraction.
• The compositional operator destroyed its operands.

Accomplishments that we're proud of

• We killed our own feature early — the fancy, demoable one. Auto-synthesis was the showpiece: seed a question, watch cards collide and fuse on their own. It demos beautifully. The evaluation said it was the noise.

What we learned

• Model diversity is an audit tool.
• Abstraction is discovered, not manufactured.
• The loop is the limit.

What's next for Altitude

• Evals. A standing quality-eval harness, not a one-off audit.
• The verification runtime — tool-using agents that fetch receipts themselves, so evidence climbs without a human courier.
• Voice at altitude — the field is already voice-operable over MCP; next is navigating by altitude ("take me to the frames," "what's stale at the ground?").

Built With

• braintrust
• claude
• codex
• elevenlabs
• fireworks
• python