Voice in real rooms
Wake-word and push-to-talk flows account for background media, silence, device changes, and natural phrasing. Owner-voice enrollment adds a practical identity layer without pretending voice alone is a security key.
Paxton Raithel / Applied AI systems
Software that earns the right to act.
I design and build applied AI systems for the conditions that polished demos tend to avoid: noisy inputs, incomplete context, interrupted workflows, constrained hardware, and actions that need a clear human boundary.
01 / Jarvis — desktop intelligence02 / Treadway — daily systemsGillette, Wyoming / 2026
Selected systems / 01—02
Built from the failure states outward.
These are living products with real boundaries, actual users, and repeatable verification—not concept renders presented as finished software.
01
Local-first macOS voice agent
Jarvis
A local-first macOS voice agent that listens, reasons, sees authorized screen context, remembers useful information, and carries out bounded computer actions without confusing a proposal with a result.
The presence system
An interface that makes internal state legible.
The orb is rendered by the native Swift app—not a portfolio illustration. Color, orbital motion, signal energy, labels, and expansion communicate whether Jarvis is listening, reasoning, observing the screen, responding, or rejecting an unverified voice.
- Listening
- Audio is open and waiting for the end of speech.
- Reasoning
- A request is being interpreted or routed.
- Observing
- Authorized visual context is being captured.
- Responding
- Jarvis is returning a spoken or visible result.
What it does
A working agent, not a voice-shaped command menu.
Screen-aware workflows
ScreenCaptureKit provides explicit visual context. Jarvis can inspect one open app or capture TradingView across 5m, 15m, 1h, and 4h, then restore the chart to its analysis default.
Bounded execution
Actions are separated from model suggestions. Consequential typing, clicking, sending, settings changes, and external effects stay behind visible confirmation and verification steps.
Context that persists
Structured local memory and retrieval carry preferences, workflows, and strategy context between sessions while keeping private mode and editable boundaries available to the owner.
Native coexistence
Jarvis lives in the macOS menu bar, yields the microphone for the separate Siri wake word, prefers the selected Razer input when present, and recovers from sleep, interruption, and permission changes.
Cost-aware intelligence
Deterministic commands stay local and fast; heavier reasoning is loaded or routed only when the task warrants it. The architecture is designed around low idle overhead and visible health state.
System architecture
Perception, policy, action, proof.
Each layer has a distinct job. A model can interpret intent or propose a plan; deterministic code owns permissions, confirmations, execution, cleanup, and whether success can be claimed.
- 01PerceiveVoice · screen · app state
- 02InterpretIntent · context · memory
- 03GatePolicy · risk · confirmation
- 04ExecuteNative action · bounded workflow
- 05VerifyPostcondition · cleanup · report
02
Private development build
Treadway
A private, installable daily-discipline product that turns recurring commitments into a calm daily path, protects recovery days, and produces an explainable weekly brief without silently sharing private reflection.
Actively developed and used on real devices. Access is intentionally limited; Treadway is not currently available as a public consumer service.
Product interface / real build
A calmer way to keep promises to yourself.
The interface favors orientation over gamification: one daily path, a clear next marker, honest recovery states, and just enough history to make patterns visible. These screens use a synthetic portfolio profile and contain no production account data.
Installable PWAOffline capableResponsive / accessible
Product behavior
The function extends well beyond the screens.
Daily path
Recurring and one-time markers, custom order, priorities, progress, hydration, reminders, and a deliberate end-of-day Close ritual.
Honest streak logic
Mountain-Time calendar rules, rest days, saved days, history, and an earned Cheat Day that never masquerades as another disciplined day.
Explainable weekly brief
A deterministic seven-day evidence window produces confidence-labeled observations, provenance, and a reviewed prompt for any AI chat.
Offline resilience
Optimistic local updates, a durable write queue, server reconciliation, and installable PWA behavior keep the path usable through unreliable connections.
Narrow accountability
A partner can receive an aggregate daily state or a fixed encouragement—never task details, journal text, or an open-ended surveillance feed.
User-owned data
Row Level Security, explicit private-text opt-in, export, account deletion, quiet hours, reduced motion, and no advertising or analytics layer.
How it is coded
A small client with deliberate infrastructure.
The deployed app is intentionally framework-free: a compact vanilla JavaScript state model, deterministic recurrence and date logic, scoped CSS, and Web Animations API effects. Supabase supplies authentication, owner-scoped Postgres Row Level Security, realtime sync, Edge Functions, and scheduled web push.
Offline changes enter a durable local outbox and reconcile with server state after connectivity returns. The weekly context engine is also deterministic: evidence provenance, confidence, prompt version, and private-text boundaries are inspectable before anything is copied to an AI conversation.
Working principles
Restraint is part of the engineering.
Evidence before claims
Automated verification and hardware-dependent behavior are reported separately. Unknowns stay visible.
Models propose; systems decide
Deterministic policy and human confirmation govern consequential actions, not model confidence alone.
Privacy is architecture
Local processing, data minimization, scoped access, and explicit consent are part of the system design.
Failure improves the product
A flaky permission, noisy room, stale cache, or partial action becomes a reproducible case and durable behavior.
Profile
Product judgment, carried through implementation.
My work sits between product judgment and implementation. I define the behavior, privacy model, failure boundaries, and acceptance criteria, then carry the system through code, testing, hardware checks, documentation, and release. The result is software shaped by repeated use—not a single happy-path recording.
Languages
Python · Swift · JavaScript · SQL
AI systems
Agent orchestration · Retrieval · Speech · Model routing · Evaluation
Platforms
macOS frameworks · MLX · Core ML · Supabase · Postgres · PWAs
Quality
Automated testing · Failure analysis · Privacy engineering · Documentation
Education
Associate Degree, Finance / Technology
Gillette College · 2026 · 3.6 GPA
Contact / Gillette, Wyoming
Selected conversations welcome.
I’m interested in applied AI, product engineering, implementation, and evaluation work where reliability and clear judgment matter.
pb8trade@gmail.com