AI Stack

Alphio AI combines a finance-native model (DMind) with general-purpose LLMs (Claude, GPT) for research, analysis, and execution.

---

1. DMind model

DMind is Alphio's finance-native LLM family — purpose-trained on financial corpora rather than retrofitted from a general model. It handles the parts of an Alphio AI query where domain grounding matters more than open-ended generation:

• Financial data structuring — turning raw market data, filings, and on-chain events into structured features the rest of the stack can reason over.
• Financial reasoning — earnings interpretation, fundamental analysis, derivative pricing intuition, scenario walk-throughs.
• On-chain data understanding — wallet behavior, token flows, contract interactions, DEX activity context.

DMind powers most of the "what does this mean" steps inside research, Market Brief generation, and Agentic Trading plan reasoning.

---

2. General-purpose LLMs

Alphio also routes to general-purpose LLMs (Claude, GPT, and similar frontier models) where they outperform a finance-specialized model:

• Long-form generation — writing the prose summary of a research run, narrating an Agentic Trading plan, drafting alert messages.
• Multi-modal inputs — image understanding (charts, screenshots), document parsing, OCR-adjacent tasks.
• Edge cases — questions that drift outside finance (general news context, regulatory text interpretation, code) where a generalist is the right tool.

These models are accessed through provider APIs and never see raw account credentials or order-placement authority.

---

3. Hybrid orchestration

Alphio AI is not a single model — it's an orchestration layer that routes each query (or each sub-step of a query) to the right model.

A typical research request flows like this:

1. Intent parsing — a small fast model classifies your request and picks a workflow.
2. Data fetch — services pull live market, on-chain, and news data.
3. Reasoning — DMind structures and analyzes the data.
4. Composition — a general-purpose LLM writes the human-readable answer.
5. Action proposal — if execution is implied (e.g. Agentic Trading), the planner constructs an approval-gated trade plan.

Each step picks its model based on the task profile. Latency-sensitive paths (chat turns) lean on smaller, faster models; depth-sensitive paths (deep research, plan generation) use bigger ones.

---

4. Privacy

What stays on Alphio infrastructure vs what leaves to a provider:

• On Alphio infrastructure — your account, broker / wallet credentials, order history, watchlist, position state, alert configurations, Credits balance and ledger. None of this is sent to external LLM providers.
• Sent to LLM providers — the textual content of your queries and the prompt context Alphio assembles (market data summaries, public filings, news snippets, and your message). Provider-side data handling follows each provider's commercial agreement.
• Never sent — payment instrument details (handled by the PSP), private keys (held by your wallet provider or custodian), and raw broker credentials (held by SnapTrade).

Trade execution itself never goes through an external LLM. The model proposes a plan; Alphio's own services place the order on the relevant rail (broker via SnapTrade, Hyperliquid on-chain, or Polymarket CLOB) after your approval.

---

warning

Exact model versions in production (DMind generation, which Claude / GPT tier per workflow), training data partners, and per-provider data retention windows are being finalized. Treat this page as a structural overview — for specific commitments, refer to the latest privacy policy or talk to the product team.