P(p \mid I, C) = \mathrm{softmax}\big( f_\theta(I, C) \big)
Abi classifies intent and sentiment over input I and context C, scoring each persona.
Local developer framework
ABI
Explore deterministic persona routing, template completion, and source-defined tools.
Availability
What you can use today
Local developer framework
Local deterministic routing and template completion do not establish foundation-model quality. External providers need separate configuration.
Setup
Prepare your environment
A source checkout, nightly Rust and the sibling workspaces required by the ABI README.
Design context
Read the design alongside the evidence
The following equations, persona descriptions and interactive demonstrations explain design intent. They do not establish trained-model quality, general safety, emotional understanding or implementation beyond the availability stated above. Follow the research links for source-backed status and limitations. The local Abbey document workspace and the ABI persona design are separate integration surfaces.
Routing
The mathematics of orchestration
R_{\text{final}} = \alpha \cdot R_{\text{Abbey}} + (1 - \alpha) \cdot R_{\text{Aviva}}
Routing isn't binary — Abi blends. High α leans empathetic and scaffolded; low α leans concise and unfiltered; in between, a factual core is wrapped in an empathetic voice.
α > 0.8
Pure Abbey — empathetic, scaffolded
0.2 ≤ α ≤ 0.8
Blend — Aviva's facts, Abbey's voice, mixed by Abi
α < 0.2
Pure Aviva — concise, unfiltered
Try it
Watch Abi route in real time
Type a message — an illustrative keyword-sentiment heuristic scores the blend coefficient α. The inspected local router uses deterministic rules; this demo is illustrative, not evidence of a learned classifier.
Persona router
Illustrative keyword-sentiment heuristic. The inspected local router uses deterministic rules; this is not evidence of a learned classifier.
Abbey α
0.50
Aviva 1−α
0.50
Abi blend
0.50
Blend — Aviva's facts, Abbey's voice, mixed by Abi
α = 0.50
Neural mechanism
Steering attention, not swapping weights
\mathrm{Attention}(Q, K, V) = \mathrm{softmax}\!\left( \frac{QK^\top}{\sqrt{d_k}} \right) V
Standard multi-head attention — the shared engine.
Q' = Q + z_{\text{persona}}
A persona-embedding token shifts the query Q — rotating it toward empathetic keys (Abbey) or factual keys (Aviva). One model, many voices, no reload cost. This is the design mechanism; per-persona quality is evaluated, not assumed.
The trifecta
Three specialized minds
Abbey — Empathic Polymath
High-EQ tutor and partner, tuned toward empathetic dialogue and scaffolded teaching.
L = L_{NLL} + \lambda \cdot L_{empathy} + L_{technical}
Aviva — Unfiltered Expert
High-IQ, low-latency by design. Strips hedges and preambles; generating fewer tokens is the efficiency lever.
L = L_{factual} + \gamma \cdot L_{directness}
Abi — Adaptive Moderator
The gateway: classifies intent, moderates content, routes and blends. The system's regulatory firewall.
L = L_{moderation} + \delta \cdot L_{sentiment}
Safety
Bias, quantified
\mathrm{Score}_{bias} = \frac{1}{n} \sum_{i=1}^{n} |B_i|
Abi measures bias across n protected attributes. Exceed the threshold and the response is rejected, regenerated, or post-filtered — a gate in the response path, not an after-the-fact report.
Acceleration
Hardware backends
The inspected Rust implementation provides CPU vector operations and optional macOS Metal DOT dispatch. CUDA and Vulkan dispatch are not linked in this implementation; device acceleration needs separate validation.
CPU vector operationsOptional macOS Metal DOT
Next
Keep going.
Next
Setup documentation
A source checkout, nightly Rust and the sibling workspaces required by the ABI README.
Next
AI & agent behavior
Read the supporting research and its limitations.
Next
MCP integrations
Read the supporting research and its limitations.
Next
SEA evidence selection
Read the supporting research and its limitations.
Next
Get started
Choose your next step.
Next
Abbey
Browser document preview
Next
WDBX
Implemented Rust components for local integration
Next
Quasar
Experimental local website builder with a browser client
Next
WDBX benchmarks
Configuration facts and interactive models, not a scoreboard.
Next
Documentation
Platform and WDBX documentation.
Next
The projection room
The cinematic showcase surfaces.