AVRAI / Software & contextual intelligence
I’m building an intelligence system for how people and the world around them fit together. Its first application helps people find places and experiences they might otherwise miss.
Start here
That last step matters to me. A useful system needs a way to hear that it has missed something.
This recording uses AVRAI Experiment’s existing SwiftUI interface and view model. Its calls reach the existing recommendation runtime through a local test harness. The scenario starts with generated onboarding preferences and routine-history fixtures; the backend is simulated. The recording shows software behavior under those inputs, not a customer visit or a claim that a real place’s atmosphere was measured.
Start with a need and time of day. Ask for a Birmingham place. Review the suggestion, then use “Wrong door” to record that it does not fit this moment. Ask again to see the next option. Recording a rejection excludes that resolved option from the next selection in this flow; it is not, by itself, evidence that a model has retrained. The visible suggestions use the current baseline selection policy; HGCNN remains a separate shadow analysis.
The beginning
I began AVRAI in March 2025 because months of curating social feeds and saving places still left me struggling to decide where to go. Sometimes I forgot those lists existed. Sometimes I opened them and still couldn’t choose. I wanted that effort to lead to experiences I could actually enjoy.
Early use has already helped my family make decisions about places and meals with less stress. That is a small beginning, but it gives the larger idea a purpose: helping people find what they didn’t know they were missing, or recognize something familiar somewhere new.
How I’m building it
I’m developing a proprietary tech stack to model “vibe”: how people, places, and situations fit together. It connects evidence across those relationships, evaluates possible futures, and uses outcomes to guide what it learns next. The goal is for knowledge gained in one context to help in another through relevant connections, with consent.
Its active state matters to me: learning from outcomes, recognizing uncertainty, and revising its understanding as lives change. I lead the product and software development across the native iOS prototype, shared runtime, contextual-model research, and feedback controls.
Someone can like a quiet café and dislike one that feels empty. The taste-signature code represents those separately. In an automated test, two candidates have equal quiet scores; adding an empty-atmosphere facet and a correction against it makes the calm candidate rank higher.
This checks how the code handles a supplied distinction. It gives me something specific to test with people.
My QETS research adapts a pretrained language backbone into an 11-layer system with a 128-dimensional context representation and a custom projection and scoring head. It exceeded 94% accuracy in HellaSwag development testing: selecting a plausible next event from four possible endings.
The question behind that work is how context can help a model understand what is likely to happen next.
The wider system
The existing AVRAI demo library explores the larger ambition through interactive scenarios and research visualizers. These use controlled data and show their evidence boundaries within each demonstration.
Turn a real-life moment into options, coordinate a plan, and explore the feedback afterward.
Describe a business, review its context, and examine proposed actions before approving them.
Follow a phrase through language interpretation, contextual weights, candidate ranking, and feedback.
What comes next
My next step is to extend private testing to five people outside my family. I want to understand what they choose, what they would change, and whether they come back when they need to decide again.
I’m looking for collaborators and pilot partners who can challenge the work and help it become useful to more people. Get in touch.