Remembers the person
Combines local session context with long-term database memory. Sudogary can retain preferences and behaviors—from nicknames to always preparing recipes for two.
Independent project · Active development
PERSONAL AI / CASE STUDY
Meet Sudogary.
A modular personal AI platform with an avatar, persistent memory, specialized services, and secure home automation—built to be useful without making advanced AI expensive. The custom wake-word model runs locally in the DJ booth; other computers use a button and push-to-talk.
.png)
01One central conversational brain
02Expandable specialist services
03Advanced models only when needed
01 See it work
From spoken request
to visible action.
These are unfiltered demonstrations of the working system: one shows Sudogary orchestrating the physical DJ environment, while the other captures him analyzing a real Mixcloud set in conversation.
“Hey Sudogary, DJ Mode on.”
A wake word and one request activate the booth’s coordinated equipment, lighting, displays, audio, and cooling.
Analyzing a real music set
Mid-conversation, Sudogary evaluates pacing, BPM interpretation, artists, subgenre positioning, tags, and Mixcloud potential.
COST-AWARE BY DESIGN
Measured usage.
Cost-aware architecture.
Sudogary uses compact context, efficient models for routine conversation, and selective escalation when a task requires deeper reasoning.
Tokens processed2.66M2,659,514 actual tracked tokens
Total tracked cost$0.30$0.297920 in the live application
Latest shown session56.3K56,346 tokens for $0.007644
Usage snapshot from the working Sudogary interface. API pricing and model selection can change over time.
02 The idea
One assistant.
Many skills.
Built around a person.
Sudogary began with a practical question: what if a personal assistant could understand not only a request, but the person making it—and reach the right private tool without forcing every interaction through the most expensive model?
The result is a chatbot with interchangeable visual and voice “skins,” a central orchestration layer called ChatCentralCore, and a growing collection of focused services. A chef can look and sound completely different from Sudogary while using the same underlying intelligence.
Routine requests stay lightweight. Deeper reasoning is available when the conversation actually needs it.
03 What it does
Works with visual context
Images can be pasted into a conversation, selected again later, and referenced as part of future questions within the session.
Controls the physical world
Connects conversational intent to Home Assistant so spoken requests can operate coordinated groups of devices and report completion verbally.
Understands local conditions
Retrieves detailed weather information with local conditions as the default, while still allowing other locations when requested.
Uses a personal recipe library
Searches Gary’s own recipe collection and applies remembered household preferences instead of treating every cooking request as a blank slate.
Builds intelligent DJ sets
Uses Mixcloud’s public API and personal music resources to suggest tracks, then helps evaluate key, vibe, energy, and sequencing.
04 Recent capabilities
New skills.
Clear status.
Sudogary is an active personal system. These updates show what is running now, what is taking shape, and what remains planned.
LIVE ON DEVICE
Custom wake-word ML
A custom “Hey Sudogary” detector, trained in Jupyter and exported to TensorFlow Lite, has been running on the Raspberry Pi for over a week. A USB microphone and speakers connect directly to the Pi. Wake-word access is confined to the DJ booth; other computers use a button and push-to-talk.
LIVE IN THE UI
Avatar activity states
The avatar can show a thinking and searching expression while Sudogary looks up information, giving the interface a visible cue that work is underway.
AVAILABLE NOW
Document interaction
The current chat experience supports document interaction. A dedicated document service is planned as a separate specialist capability.
DEDICATED SERVICE · UPCOMINGIN USE
Structured tool results
Service results can be presented as organized, copyable tables. The current interface demonstrates this with detailed track-level music analysis.
IN DEVELOPMENT
Local voice options
Gary is evaluating local voices as a possible replacement for the current hosted speech response, with the goal of reducing external voice-model calls. Speech-to-text still uses a transcription service; live transcription is being considered separately.
HUMAN-GUIDED
Multi-step DJ workflows
Sudogary can search music sources, suggest tracks, analyze musical attributes, and help shape a set. Gary reviews and chooses the suggestions, keeping the workflow collaborative.
05 In action
VOICE-TO-ENVIRONMENT WORKFLOW
“Hey Sudogary…
DJ Mode on.”
01
Wake
The custom wake-word model runs locally on the Raspberry Pi and recognizes “Hey Sudogary” in the DJ booth, then gives a confirmation signal. Other computers use a button and push-to-talk.
02
Interpret
ChatCentralCore understands “DJ Mode on” and selects the Home Assistant service.
03
Orchestrate
The infinity mirror, soundbar, stereo, two TVs, cooling fan, and subwoofer turn on as one coordinated scene.
04
Confirm
Sudogary responds verbally when the environment is ready—even if the visual interface was never opened.
06 Applied AI architecture
Applied AI,
connected to action.
Sudogary combines a conversational model with conventional software services, memory and retrieval, a locally running wake-word model, and physical automation. Each part has a defined role; the conversational model is only one part of the system.
EXPERIENCEDJ boothUSB microphone · USB speakersWake-word activation
EXPERIENCEOther computersChat interfaceButton + push-to-talk
BOOTH AUDIOCHAT INPUT
01 · MACHINE LEARNING + EDGE INFERENCERUNNING LOCALLY
Custom wake-word modelTrained in Jupyter · TensorFlow Lite exportInference on Raspberry Pi · booth-scoped microphone
WAKE EVENT / USER MESSAGE
02 · CONVERSATIONAL AICLOUD MODELS · COST-AWARE ROUTING
ChatCentralCoreStreaming conversation · model selection · context assembly
Memory & retrievalSession context · long-term memory · selected images and documents
Tool useCategory routing · service APIs · structured results
Agentic workflowsMulti-step tasks with user review and control
TOOL CALLS
03 · CUSTOM SERVICESINDEPENDENT APPS + APIS
Weather
Recipes
Mixcloud
Personal music
Document service upcoming
Home Assistant
DEVICE COMMANDS
04 · PHYSICAL AUTOMATIONHOME ASSISTANT
Infinity mirror
Soundbar + stereo
Displays
Cooling fan
Subwoofer
Designed as a system, not a model demo
Cost-aware model routing handles conversation; specialist code handles retrieval, APIs, and device commands. The same assistant can answer a question, return structured results, or request an authorized physical action through Home Assistant.
07 The signature skill
The ultimate
DJ assistant.
Sudogary’s defining capability is not simply recommending music. He can research a requested genre, scan Gary’s public Mixcloud presence, draw suggestions from preferred music resources, and help shape the result into a playable journey.
DISCOVERANALYZESEQUENCE
Each candidate can be examined for musical key, vibe, energy, and its role in the set. The assistant remembers Gary’s taste and working preferences, turning track discovery into a collaborative process rather than a generic list.
08 Private by design
Connected without
giving everything away.
Sudogary has an online presence with authenticated sessions and automatic logout. An encrypted VPN connection links the DJ computer to the Raspberry Pi hosting ChatCentralCore and its services.
SECURE PATH
Authenticated interfaceEncrypted private linkLocal services & data
Selected databases remain on Gary’s own network while still being available to the assistant when authorized.09 Built by Gary
Designed as a system.
Proven piece by piece.
Gary designed the product, architecture, capabilities, service boundaries, behaviors, and experience. Codex served as a coding partner throughout implementation.
The code was checked and measured at each stage, with working services added incrementally as real needs emerged. Sudogary is not a packaged chatbot with a new name; it is a personal software platform assembled around a deliberate architecture.
10 Why this project matters
Applied AI
has to work in context.
Sudogary demonstrates the work between a capable model and a useful system: designing interfaces, choosing when to use AI, connecting APIs, retrieving the right personal context, shaping reliable tool results, and linking digital decisions to physical devices.
It brings conversational AI, custom machine learning, local edge inference, modular services, persistent memory, and home automation into one evolving project. It also makes the tradeoffs visible: what runs on the Raspberry Pi, what uses external models, when the user stays in control, and where new services fit.
The goal is a personal assistant that grows through real use—one carefully integrated capability at a time.
GARY BLANCHARD / SELECTED PROJECT