Sig AI Dialog Agents
Agent Architecture
We use small language models trained on narrow domains or tasks. Each task is defined by the domain "ontology" which defines the language "slots" for NLU (understanding user intent)
The generation of dialog is optionally done using large language models, but typically using deterministic code relying on corporate policies (e.g. pricing or capacity constraints), session state and system state
Session state comes from in-memory or stored context. System state and corporate business policies come from database queries, API calls. This results in predictable and "explainable" AI
Models are trained using supervised fine-tuning (SFT) on each domain using your data or synthetic data generated from ontology
Sig AI Dialog Agents Advantage
Real-time or batch-mode (chats or email interaction)
local, on-prem, or private cloud deployment
Open-source, free and private language models (Llama 3,x 8B ,T5 Large or XL etc.)
Trainable on single-GPUs A100
Inference on just CPU servers, no GPUs needed
No "token" usage counting etc.
Continuous learning from incremental fine-tuning
Typical use cases
Multi-turn dialogs between users and systems for inquiries and transactions
Multi-turn dialogs between users and systems for negotiating prices or other terms
Multi-turn dialogs between users and systems for negotiating change (e.g. changes to previously agreed terms or service)
Automate "one-to-many" interactions: e.g. request for proposals from multiple service providers
Business roles that can be played by the Equate AI Dialog agent include: sales agent, Finance AP agent, buyers/planners, procurement, Finance AR agent, Customer service, Customer onboarding etc.
