AI Copilot Integration

FAQ #

• What types of SaaS products can you add a copilot to?

  Any web application with a defined action surface — maps, dashboards, CRMs, project tools, IoT monitoring. The key requirement is a frontend with 10+ distinct user actions. We work primarily with React frontends and REST or WebSocket backends, but the architecture adapts to your stack.
• What is a “command contract” and why does it matter?

  The command contract is the typed schema defining exactly which actions the AI can execute — {type: "load_dataset", params: {...}}, for example. A well-designed contract stops the AI from hallucinating invalid actions. We design the schema before any implementation code is written.
• How long does it take to add a copilot to an existing product?

  The five-day Diagnostic Architecture Audit delivers the command schema, backend architecture, and integration plan. A first working copilot on a core flow follows in 10–15 days; the full rollout depends on your command surface.
• Does the integration replace the existing UI?

  No. We deliver a standalone chat panel component that integrates alongside your existing frontend. Your product’s UI stays untouched — the copilot is a new control surface, not a replacement. Users choose how to interact.
• Which AI model powers the copilot?

  Anthropic Claude (currently Haiku for speed and cost) by default. The backend proxy is model-agnostic — GPT-4o, Gemini, or your preferred provider are config changes, not rewrites. We scope the model during the Diagnostic Architecture Audit.
• How is the AI API key protected?

  The key never touches the browser. All AI calls go through a Rust/Axum backend proxy on your infrastructure. The browser sends a plain message; the backend adds the key, calls the AI, and returns typed commands. No dangerouslyAllowBrowser, no key exposure.
• What happens when the AI returns an unexpected or malformed response?

  We build a lenient parse pipeline that handles markdown-fenced JSON, partial responses, and error envelopes — the formats AI models actually return under load. Unknown or malformed commands are dropped with a warning log, not a user-visible crash.
• Is abuse prevention included?

  Standard on every engagement: multi-window rate limits (per-IP, per-session, global — minute/hour/day/month), token budget enforcement, request size caps, and Postgres-persisted counters.
• Can the copilot be scoped to individual users or sessions?

  Yes — scoped during the Diagnostic Architecture Audit. Isolated per-user sessions require auth; a shared history view suits multi-operator dashboards. The architecture supports both without a rewrite.
• Do you sign NDAs and formal contracts?

  Yes. NDA before any codebase access. Signed contracts with fixed scope per phase. Wyoming LLC — you’re contracting with a registered entity, not an individual.