Assistant
Chat and voice
A 7-stage pipeline with dual-model strategic planning. Full-duplex voice through Hume EVI with intelligence parity to the text interface.
Chat pipeline
7-stage processing
1
Validate request
Authentication, workspace resolution, rate limiting.
2
Resolve conversation
Load or create session. Retrieve message history.
3
Classify and route
Complexity assessment. Strategic intent detection. Tool pre-selection.
4
Load context
7 parallel sources: brain, memory, skills, meetings, tasks, preferences, signals.
5
Build system prompt
Compose skills, inject strategic directive, entity context, tool instructions.
6
Window messages
Token management. Context window optimization. Message pruning.
7
Execute stream
Tool loops with streaming response. Turn tracking. Cost logging.
Strategy
Dual-model planner
Complex requests trigger a dual-model pipeline. Gemini ingests massive context cheaply to produce a strategic directive. Claude executes with nuanced tool use.
Message Classifier
detects: strategic_intent
|
Gemini 3.1 Pro (STANDARD, 1M context)
Ingests: entity profiles, comms history, brain, tasks, calendar
Produces: StrategicDirective
user_intent
complexity_assessment
strategic_analysis
entity_context_summaries
recommended_tool_sequence
communication_guidance
|
XML injection into system prompt
|
Claude Sonnet 4.6 (STRONG)
Executes tools guided by directive
Streams responseTools
200+ tools across 18 domains
5-gate selection: complexity classification, intent matching, skill profiles, cap enforcement, planner boosting.
| Domain | Tools |
|---|---|
| Intelligence | 24 |
| CRM | 22 |
| 18 | |
| Meetings | 16 |
| Tasks | 14 |
| Contacts | 12 |
| Workspace | 10 |
| Documents | 8 |
| Admin | 8 |
| Social | 6 |
Voice
Hume EVI voice interface
Full-duplex voice with intelligence parity to chat. Same context, same tools, same brain. Three-state floating widget.
User speaks
|
Hume EVI WebSocket
|
CLM Endpoint (Next.js API)
|
Voice Context Loader (7 sources, 15-min cache)
|
Anthropic + 19 tools (max 5 loops)
|
SSE Stream
|
Hume TTS
|
User hears19 voice tools
create_tasksearch_tasksupdate_taskget_todays_meetingsschedule_meetingreschedule_meetingfind_availabilitysearch_meetingssearch_contactsget_contact_profileupdate_contactdraft_emailget_email_threadget_briefingget_vibe_cardsget_suggestionsstore_memoryrecall_memorystart_contact_warmup
Context
7 parallel context sources
Workspace Brain
6 pinned blocks with synthesized workspace intelligence. 8,800 tokens.
Living Brain
Pinned memories and relevant episodic memories via semantic search.
User Skills
Semantic match against user-defined skill profiles. Top-3 composed.
Today's meetings
Upcoming meetings with participant context for scheduling awareness.
Active tasks
High-priority in-progress tasks for status awareness.
AI preferences
User-configured behavior preferences and communication style.
0Pipeline Stages
0+Tools
0Domains
0Voice Tools
Talk to Sentigen
Have questions about what you're reading? The intelligence behind this platform can walk you through it.