Nineteen crates.
Four layers.
Engineered for the regulator-defensibility constraint from crate one — type-safe, memory-safe, async-first. A hexagonal workspace where dependencies point down and the domain crate depends on nothing. Explore it on the right.
HTTP/gRPC server, auth, routing — the composition root.
Live map · hover a crate to see its dependencies light up — the rule points down.
Production AI is too important for Python in production.
You’ve watched a team retrofit a Python ML stack into production. The ergonomics win gets eaten by the production cost, every time. DEINO is Rust from crate one — and the reasons are load-bearing, not aesthetic.
Type & memory safety
No GIL contention, no runtime errors that should have been compile errors, no memory leaks under sustained load. The cost Python extracts in production, paid once at compile time.
Determinism
In regulated AI, “almost reproducible” and “bit-for-bit reproducible” are the difference between passing audit and not. Rust’s guarantees are why DEINO can promise the latter.
Manifests by typed code
Audit manifests are generated by the type system, not bolted on. Calibration regressions are enforced by types. The compiler is part of the compliance story.
Tokio async
A single async runtime across the whole stack — streaming, A2A orchestration, ensemble fan-out. P95 < 150ms on million-node ontologies.
Reasoning, orchestration, grounding.
The combination — sister-LLMs for reasoning, A2A for orchestration, the GNL for grounding — is what produces defensible forecasts where a single LLM would hallucinate.
5 ensemble sister-LLMs
Five independent LLM clients run in ensemble, each with separately-tuned prompts and calibration. deino-oracle synthesizes them into one distribution with explicit confidence. Any provider — you bring the keys.
25 specialized A2A agents
The agent-to-agent layer orchestrates 25 agents for distinct sub-tasks: disambiguation, axis calculation, network traversal, calibration weighting, manifest construction. Each versioned, each logged to the manifest.
The whole stack is what your senior engineers would have built with fifteen years and a clean sheet of paper.
DEINO engineering principle · architecture.mdFrequently asked questions.
Why Rust, how the workspace is organized, and what the sister-LLMs and A2A agents actually do.
Why is DEINO built in Rust instead of Python?
Rust gives DEINO type and memory safety, a single Tokio async runtime, and — most importantly for regulated AI — bit-for-bit determinism. In an audit context, the difference between "almost reproducible" and "bit-for-bit reproducible" is the difference between passing and failing. Audit manifests are generated by the type system rather than bolted on afterward.
How is the architecture organized?
DEINO is a hexagonal Rust workspace where dependencies point down and the domain crate depends on nothing. It combines five ensemble sister-LLMs for reasoning, 25 specialized agent-to-agent (A2A) agents for orchestration, and a versioned four-layer Geo-Neural Layer for grounding — the combination that produces defensible forecasts where a single LLM would hallucinate.
What do the sister-LLMs and A2A agents do?
Five independent LLM clients run in ensemble, each with separately-tuned prompts and calibration; deino-oracle synthesizes them into one distribution with explicit confidence. The 25 A2A agents handle distinct sub-tasks — disambiguation, axis calculation, network traversal, calibration weighting, and manifest construction — each versioned and each logged to the audit manifest.
Architecture deep-dive with Carlos.
Bring your platform team. A 90-minute whiteboarding session on crate boundaries, deployment topology, and your specific integration points.