← All projects
NCMA / PERSONALAGENT / LOCAL-FIRST INTELLIGENCE
A voice. A presence. A system behind it.
A personal intelligence system joining voice, memory, knowledge, MCP tools and web research, with an orb that makes the state of the voice interaction visible.
MY CONTRIBUTION
Independent system architecture and integration across presentation, runtime boundaries, voice interaction, resource-aware inference and guarded tools.
TOOLS & METHODS
Local + cloud inference · wake word · STT / TTS · MCP · memory · knowledge retrieval · web research
CHALLENGE
A personal assistant has to work across voice, tools and knowledge while sharing local compute with other workloads. Its interface must remain responsive without confusing presentation with execution authority.
DESIGN DECISIONS
Separate the presentation contract from inference and tool execution. Validate presentation events against session identity and sequence; drive the orb from speech amplitude. Treat local or cloud inference as a resource-aware routing decision.
CONSTRAINTS
Voice, knowledge and tools have different latency and resource needs. Local/cloud routing and session identity must remain explicit; a visible speech state cannot grant execution authority.
Idle / original Jarvis UI
Silent visual demonstration · runtime not connected
Editable voice states · sample speech envelope · microphone off
The workflow, step by step.
01 / Voice & intent
Wake word, speech input and speech output form the voice-facing interaction.
02 / Presentation contract
Session and sequence boundaries govern visible state. The speech envelope drives the orb without repainting the entire interface.
03 / Runtime routing
Local and cloud inference account for the task and available GPU and memory resources.
04 / Context & tools
Memory, knowledge, research and MCP tools supply context behind explicit runtime boundaries.
05 / Execution authority
Visible state is not permission to run a tool. Guarded actions remain subject to the relevant execution contract.
IMPLEMENTATION & VALIDATION
Wake word, STT/TTS, memory, MCP tools and web research form the interaction context. The presentation contract uses session identity and sequence; speech amplitude drives the orb within its visible states.
RESULT & EVIDENCE
The reviewed interface and presentation contract establish distinct idle, listening and speaking states. This Figma case reuses the editable orb and its demonstration states.
EVIDENCE LIMIT
The portfolio demonstrates a sample speech envelope. It connects no microphone, live model, desktop action or tool runtime. End-to-end physical acceptance remains separate from automated or structural checks.
IDLE / LISTENING / SPEAKING
At rest, the orb is calm. Listening and speaking use distinct motion amplitudes so the interaction state remains legible.
RESOURCE TRADE-OFF
Local models share GPU and memory with other workloads. Hybrid routing preserves a path for tasks that should run remotely.
TRADE-OFFS & LESSONS
Local-first execution adds resource constraints; cloud routing adds a separate boundary. The Figma interaction demonstrates presentation states with sample envelopes. Microphone operation and physical end-to-end runtime acceptance require their own evidence.
Explore architecture ↗
Next case: Knowledge Runtime / 26AI ↗
About the work and career ↗
VALTER LOURENÇO JUNIOR / NCMA SYSTEMS
← All projects
NCMA / PERSONALAGENT / LOCAL-FIRST INTELLIGENCE
A voice. A presence. A system behind it.
A personal intelligence system joining voice, memory, knowledge, MCP tools and web research, with an orb that makes the state of the voice interaction visible.
MY CONTRIBUTION
Independent system architecture and integration across presentation, runtime boundaries, voice interaction, resource-aware inference and guarded tools.
TOOLS & METHODS
Local + cloud inference · wake word · STT / TTS · MCP · memory · knowledge retrieval · web research
CHALLENGE
A personal assistant has to work across voice, tools and knowledge while sharing local compute with other workloads. Its interface must remain responsive without confusing presentation with execution authority.
DESIGN DECISIONS
Separate the presentation contract from inference and tool execution. Validate presentation events against session identity and sequence; drive the orb from speech amplitude. Treat local or cloud inference as a resource-aware routing decision.
CONSTRAINTS
Voice, knowledge and tools have different latency and resource needs. Local/cloud routing and session identity must remain explicit; a visible speech state cannot grant execution authority.
Idle / original Jarvis UI
Silent visual demonstration · runtime not connected
Editable voice states · sample speech envelope · microphone off
The workflow, step by step.
01 / Voice & intent
Wake word, speech input and speech output form the voice-facing interaction.
02 / Presentation contract
Session and sequence boundaries govern visible state. The speech envelope drives the orb without repainting the entire interface.
03 / Runtime routing
Local and cloud inference account for the task and available GPU and memory resources.
04 / Context & tools
Memory, knowledge, research and MCP tools supply context behind explicit runtime boundaries.
05 / Execution authority
Visible state is not permission to run a tool. Guarded actions remain subject to the relevant execution contract.
IMPLEMENTATION & VALIDATION
Wake word, STT/TTS, memory, MCP tools and web research form the interaction context. The presentation contract uses session identity and sequence; speech amplitude drives the orb within its visible states.
RESULT & EVIDENCE
The reviewed interface and presentation contract establish distinct idle, listening and speaking states. This Figma case reuses the editable orb and its demonstration states.
EVIDENCE LIMIT
The portfolio demonstrates a sample speech envelope. It connects no microphone, live model, desktop action or tool runtime. End-to-end physical acceptance remains separate from automated or structural checks.
IDLE / LISTENING / SPEAKING
At rest, the orb is calm. Listening and speaking use distinct motion amplitudes so the interaction state remains legible.
RESOURCE TRADE-OFF
Local models share GPU and memory with other workloads. Hybrid routing preserves a path for tasks that should run remotely.
TRADE-OFFS & LESSONS
Local-first execution adds resource constraints; cloud routing adds a separate boundary. The Figma interaction demonstrates presentation states with sample envelopes. Microphone operation and physical end-to-end runtime acceptance require their own evidence.
Explore architecture ↗
Next case: Knowledge Runtime / 26AI ↗
About the work and career ↗
VALTER LOURENÇO JUNIOR / NCMA SYSTEMS