Meet INGRID
A knowledgeable technician sitting beside you
INGRID is the reasoning layer built into BlueCrow Studio. She doesn’t just report numbers — she separates what was measured from what she’s inferring, and she says so plainly when something can’t be answered from local data alone. She’s never designed to guess with confidence she hasn’t earned.
Character Design
This is INGRID
One face, one silhouette, one voice. Auburn hair, teal eyes, a dark midnight-blue coat, and the BlueCrow raven at her shoulder — the canonical design she keeps everywhere BlueCrow presents her.
Commander & Companion
You only ever talk to INGRID
INGRID is the only intelligence you ever speak with — the interface, the personality, and the one consistent voice across everything BlueCrow does, technical or creative. You never switch assistants, never manage a provider yourself, and never see the machinery underneath.
When something needs real engineering work — a diagnosis, a technical review — INGRID delegates it to C.O.R.V.U.S., her engineering core, and brings the reviewed result back in her own words. She commands. C.O.R.V.U.S. engineers. The decision is always yours.
What INGRID can do today
System-aware conversation
INGRID receives your system’s current live telemetry and health state directly, so her answers reflect what’s actually happening on your machine right now.
Deterministic commands
A fixed set of phrases — navigation, “scan my hardware,” “how is my system,” and close variants — are recognized and answered instantly and locally, without contacting any AI model.
Thermal reasoning
Per-component CPU/GPU classification, plus cross-metric reasoning about whether a workload is CPU-bound, GPU-bound, or balanced, and whether cooling looks adequate for the load.
Live knowledge lookups
Cross-source lookups for things local telemetry can’t answer alone, such as current NVIDIA driver advisories, layered on top of INGRID’s local reasoning.
Open-ended conversation
Anything outside INGRID’s recognized commands and reasoning modules is sent to a locally running Ollama model. If Ollama isn’t reachable, she says so plainly rather than failing silently.
Voice conversation
Push-to-talk speech recognition (OpenAI Whisper, with a local Windows fallback) transcribes what you say; INGRID replies out loud in Marin, a natural OpenAI voice.
Living reactor and avatar
INGRID's on-screen reactor emblem and avatar reflect her real state — idle, listening, thinking, speaking, or creating — driven by the same live signal, not separate visual layers that can drift out of sync.
Web search, with real context
INGRID can search the web and summarize what she finds, then keep track of the search itself — its sources, and which one you're currently focused on — so she can open a specific source or answer “tell me more about the first one” against what was actually returned, not a guess.
Hands-free wake word
Say “Hey INGRID” and she starts listening, so a question doesn’t need a keyboard or a button first.
Launching and controlling your applications
Ask for an app or a game by name and INGRID opens it, brings it to the front, or closes it. She finds what’s installed from the shortcuts, registry entries and Microsoft Store packages already on your machine, rather than a list someone maintained by hand.
INGRID Link — reaching her from your phone
A paired phone can talk to INGRID over a private, encrypted link on your own network, never the public internet: two-way voice, and her live camera view staying up for the whole call.
Voice lock
Optional speaker verification, so spoken commands are acted on only when it’s you. It stays completely dormant until you enrol, and asks before it ever listens.
Follow-up questions and conversational state
She holds onto what you were just talking about and what she just told you, and recognizes a first slice of natural rephrasings of a follow-up — “give me the short version,” “go deeper on that” — beyond her originally recognized exact phrases, without ever inventing a target she isn't confident about.
What works without an internet connection
Most of INGRID’s reasoning runs entirely on your machine: system-aware conversation, deterministic commands, and thermal reasoning all work with no internet connection at all.
A few things reach outside your machine. Open-ended conversation is handled by a locally running Ollama model — still local, but a separate install. INGRID’s live knowledge lookups, like current driver advisories, need an actual internet connection to work at all. Voice conversation prefers OpenAI for recognition and speech, but falls back automatically to Windows’ own built-in speech engine when no connection or API key is configured, so talking to INGRID still works offline. As INGRID takes on new capabilities, we won’t assume they’re all offline-capable by default.
How she shows up on screen
INGRID’s presence adapts to how you’re using BlueCrow right now — live today, not a mockup.
Standard
The everyday view — INGRID present alongside your telemetry, tools, and chat.
Compact
A smaller footprint for when you want BlueCrow open but out of the way.
Focus
INGRID as the centerpiece — the conversation is the point, everything else recedes.
In Testing
INGRID is starting to see
A camera input pipeline and a new Vision panel — live device selection, a real preview feed, honest Offline/Connecting/Active/Error status, resolution, and measured frame rate. INGRID gains a matching Observing state, so she visibly attends to the feed while it’s live. Built and unit-tested, including against the real (currently camera-less) device-enumeration path — not yet verified end to end against physical camera hardware.
None of the following exist yet — they’re clearly labeled “coming soon” in the panel itself, never implied to already work:
- Object recognition
- Scene understanding
- Facial tracking
- Text recognition
- Emotion recognition
How It Works
One assistant, coordinating everything underneath
You talk to INGRID once, in chat or by voice. She decides which engine or connected provider actually handles the request, then hands back a single, consistent response instead of a patchwork of separate tools.
You
Ask a question, request art or music, or check on your system — by chat or by voice.
INGRID
Receives the request and decides what should happen next.
Chooses and coordinates the appropriate engine or provider
Returns one consistent BlueCrow experience
What INGRID coordinates
Some of this runs today against real, connected accounts. The rest is an honest look at where that coordination is headed next.
Handles open-ended conversation whenever a request falls outside INGRID's recognized commands and reasoning modules — runs entirely on your machine, no account required.
Handles voice conversation with INGRID end to end — speech recognition and spoken replies. Connect your own account in Creative Services and it can also generate Art Studio images; that route is verified against a live account and still being tested more broadly.
Generates Music Studio tracks once you connect your own account in Creative Services. Verified end to end against a live ElevenLabs account, though Music Studio itself is still in testing.
Not connected today. The long-term idea is to let conversation or creative requests route through Claude the same way Ollama and OpenAI already do.
OpenAI is the only connectable image provider right now. Adding more alongside it is a long-term goal, not active work.
Routing richer telemetry and optimization actions through the same provider model INGRID already uses for creative requests — an early idea, not yet designed or scheduled.
Where This Is Going
A richer presence, honestly not built yet
- Natural expressions, eye movement, and breathing — not just a glow that changes color.
- Gesture and posture that reads as presence, not just a state indicator.
- An always-available overlay, so INGRID isn't confined to one window.
Direction, not a promise of a date — see the roadmap for exactly what’s live versus planned versus research.
Safety & Authority
She asks before she acts
INGRID can now do things, not only describe them — open an application, run a safe cleanup, work through a task you gave her. Every one of those goes through the same gate: the action is classified by risk, you approve it, a checkpoint is taken first where one applies, and the result is reported as what actually happened rather than as success by default. Anything she could not verify, she says she could not verify.