Skip to main content

AI & ML Roadmap

Draft

This page collects proposed AI/ML enhancements for Cat Wallet. These are proposals, not commitments — scope, sequencing, and feasibility need review before any of this is built.

Cat AI Chat and Infernet already give Cat Wallet an agentic assistant and a verifiable off-chain inference layer (see Cat AI Chat and Infernet Models), plus a background Autonomous Security Sentinel. The items below extend that foundation rather than replacing it.

Trust & Verifiability​

Model explainability layer​

Infernet risk assessments influence real money decisions (e.g. "this address looks risky"), so raw verdicts aren't enough on their own.

  • Surface a confidence score alongside every risk verdict, not just a Safe/Warning/Critical label.
  • Attach a short reasoning trace (which signals drove the score — e.g. contract age, verified source, prior flagged interactions) so the "human-readable risk assessment" is inspectable, not a black box.
  • Store the trace alongside the cached scan result (see the Sentinel's 24h cache) so it can be shown again without re-running inference.

On-chain verifiable inference​

Infernet's core value proposition is verifiable off-chain compute — today Cat Wallet consumes model output but doesn't lean on that verifiability for the parts of the product where it matters most.

  • Extend on-chain verifiable inference specifically to the risk-scoring model, so a result is provably produced by the claimed model/weights, not just trusted because it came from our backend.
  • Publish the verification proof (or a reference to it) next to the risk verdict shown in-app, so a user (or an auditor) could independently confirm the score wasn't fabricated or silently swapped for a cheaper model.

Assistant Capability​

Multi-turn financial reasoning​

Cat AI Chat currently parses one intent per message and drafts a single transaction for approval (see How It Works).

  • Allow the assistant to plan multi-step actions end-to-end — e.g. "swap 0.5 RITUAL for rUSD, then stake the rUSD" — as a single conversational request.
  • Render a plan preview showing each step, its simulated outcome (slippage, price impact, expected stake yield), and dependencies between steps, before any signature is requested.
  • Preserve the existing safety model: the assistant still only prepares transactions; each step in the plan still requires its own wallet confirmation, executed in order.

Personalized financial assistant memory​

The assistant currently has no persistent memory of a user's habits across sessions.

  • Add opt-in, embeddings-based memory (e.g. a vector store keyed by wallet address) capturing patterns like typical swap pairs, usual transaction sizes, or preferred slippage tolerance.
  • Use that context to make suggestions feel personalized (e.g. flag an unusually large transfer relative to the user's own history) — this overlaps with, and can feed, the anomaly detection work below.
  • Strictly opt-in and disclosed, consistent with the existing privacy posture (private keys and signing material never leave the device or reach the model; see Security → Data handling). Users can clear this memory at any time.

Security & Fraud Detection​

These extend the Sentinel from address-reputation checks into pattern- and content-based detection.

Phishing-domain / impersonation NLP classifier​

  • Automatically scan links shared in chat with an NLP classifier trained on phishing/typosquat patterns before a link is rendered as clickable.
  • Flag likely impersonation domains (lookalike characters, suspicious TLDs, brand-name substrings) with an inline warning rather than blocking outright, keeping the user in control.

Smart contract bytecode risk classifier​

  • Train an ML classifier on bytecode patterns associated with common rug-pull/honeypot contracts (e.g. transfer restrictions that only trigger after liquidity add, hidden owner-only mint functions).
  • Run this before a user's first interaction with an unfamiliar contract, complementing — not replacing — the existing LLM-based diagnostics in Infernet, which reason more generally about contract/token data.

Anomaly detection on transaction patterns​

  • Beyond the Sentinel's current address-reputation scanning (every ~15 minutes, 24h cache per address), add a model that profiles each wallet's own behavior — typical amounts, frequencies, counterparties, times of day.
  • Flag transactions that break from that profile (e.g. a sudden large outbound transfer to a never-seen address) as a behavioral anomaly, independent of whether the destination address itself has any known reputation.
  • Deliver these alerts through the same proactive notification path the Sentinel already uses (feed + toast + bell), with the same anti-spam throttling (one alert per user/address per 7 days).

Suggested sequencing​

  1. Model explainability layer — lowest lift, directly improves trust in what already ships.
  2. Anomaly detection on transaction patterns — reuses Sentinel's delivery/caching infrastructure.
  3. Phishing-domain / impersonation NLP classifier — contained scope, clear safety win.
  4. Smart contract bytecode risk classifier — needs a labeled training set; longer lead time.
  5. Multi-turn financial reasoning — meaningful UX change to Cat AI Chat; needs careful plan-preview UX work.
  6. Personalized financial assistant memory — depends on multi-turn reasoning landing first for full value; privacy/consent UX needs design.
  7. On-chain verifiable inference — highest complexity, depends on Ritual Infernet's own verifiable-inference primitives maturing.