Classify the task
Identify whether the product needs retrieval, extraction, coaching, roleplay, generation, or a deterministic function.
AI architecture
Put product controls around the model call.
SoloFrame treats the LLM as one component inside a workflow. Retrieval, task routing, safety classification, approval gates, audit records, and measurement determine whether the result becomes useful product behavior.
Six controls
A bounded path from user intent to product action.
Apply safety policy
Run the classifier, redaction, data-handling, and product rules that belong before model inference.
Retrieve context
Fetch the minimum relevant product and user context instead of injecting the full knowledge base.
Route the model
Select a model that meets the task's quality, latency, privacy, and cost requirements.
Gate the action
Require user or operator approval before sensitive external communication or system changes.
Measure the result
Record usage, latency, safety events, output quality, and the product-specific outcome.
MAIA classifier service
Digital Wellness Academy classifies distress before coaching.
The PHI-aware vertical selects MAIA in its manifest. Commercial products select a non-clinical configuration. This keeps safety behavior explicit at product boot instead of burying it inside prompt text.
- Classifier runs before the coaching model.
- Signal records store hashes and metadata instead of raw message text.
- Model promotion requires validation against a defined baseline.
- A failure can route to conservative product behavior.
Request pathPHI-aware vertical
user message ↓ redaction and policy ↓ MAIA distress classification ↓ retrieval and model routing ↓ controlled response ↓ audit and outcome signal
Cost control
Spend model tokens where they change the product result.
Cost architecture starts with task design. Smaller prompts, focused retrieval, deterministic tools, caching, and fit-for-purpose models reduce waste before infrastructure discounts enter the picture.
Task routing
Use a stronger model for work that needs it and a lower-cost model for bounded classification or transformation.
Focused context
Retrieve the few records or lesson chunks that answer the current task.
Deterministic services
Move safety, validation, state transitions, and calculation into testable code where possible.
Cache stable input
Reuse instructions and product context that do not change between adjacent turns.
Measure by task
Attribute tokens, latency, errors, and outcome quality to the product action that caused them.
Keep providers replaceable
Route through an abstraction that lets the product change a model without rewriting the workflow.
Measured economics
The architecture pays for itself on recorded runs.
SoloFrameHub GTM OS
Coaching and roleplay sessions cost 6.2x less than a naive single-frontier-model design for the same task mix, with safety and retrieval steps included rather than skipped.
Digital Wellness Academy
PHI-aware coaching sessions run at 4.4x less than naive designs while adding MAIA classification before every model call.
Multipliers come from recorded sessions over each vertical's production task mix, not projections. The mechanics behind them are described above and in the cost architecture guide.
Direct answers
Frequently asked questions
Does SoloFrame use one AI model for every task?
No. The model registry maps tasks to fit-for-purpose models and supports per-vertical overrides. The product can route classification, coaching, extraction, generation, and embeddings through different models.
What is MAIA?
MAIA is the platform classifier service used by Digital Wellness Academy for distress routing before a coaching model sees the message. Commercial verticals can select a no-op classifier when that clinical safety path does not apply.
How does retrieval work across products?
Each vertical chooses the retrieval path that fits its corpus and operating needs. The platform supports Postgres and pgvector patterns as well as bounded product-local retrieval.
Can an AI action contact a customer or change external systems without review?
Product workflows can place approval gates before external side effects. The correct control depends on the action, buyer policy, and integration.
Start with the buyer job
Bring one expensive workflow, the buyer who owns it, and the measure that proves improvement.
SVTech will assess whether SoloFrame fits the product, compliance, data, and operating requirements before proposing a build.