Real-time AI interview assistant — invisible to your interviewer
Live transcription → instant AI answers → stealth HUD overlay
Click to watch the demo
SolveWatch listens to your interview through your microphone, transcribes questions in real time using on-device Whisper, and streams AI answers into a floating HUD overlay — completely invisible to Zoom, Google Meet, Teams, and every other screen-capture tool.
No browser extension. No cloud audio. Runs on your machine.
The HUD overlay — always on top, invisible in screenshare
Settings page — configure providers, fallback chain, and models
On-device Whisper STT with speaker diarization
| Feature | Details | |
|---|---|---|
| 🎤 | Flexible STT | Three modes: Local Whisper (MLX on Apple Silicon, openai-whisper on Windows — fully offline), OpenAI Whisper API (cloud, no local model), or Deepgram Cloud (nova-2, ~300ms latency, $0.0059/min with $200 free credit). Switch in settings, no restart needed. |
| 👁️ | Invisible overlay | setContentProtection(true) — same OS API used by banking apps. Excluded from Zoom, Meet, Teams, Loom, OBS, and all screen recording tools. Works for full-screen share, not just window share. |
| ⚡ | Sub-second answers | First token in ~200 ms with Groq. Answers stream token-by-token into the HUD while the model is still generating. |
| 🔁 | Multi-provider fallback | Configure a cascade: OpenAI → Groq → Gemini → Claude → Ollama. If one fails or rate-limits, the next kicks in automatically. |
| 🧠 | Conversation memory | Remembers the last 3–5 Q&A pairs. Follow-up questions like "what are its trade-offs?" work correctly. |
| 📸 | Screenshot analysis | Monitors a folder for new screenshots, runs OCR (Tesseract) + AI, and shows the answer in the HUD. Great for coding problems shared on screen. |
| 👤 | Speaker identification | Local-mode only: SpeechBrain ECAPA-TDNN filters your voice so only the interviewer's questions trigger AI answers. Enroll a 30s sample once. Deepgram mode uses built-in diarization instead. |
| 📊 | Grafana observability | Optional OpenTelemetry export to Grafana Cloud — AI latency, token spend, STT pipeline health, host metrics. Import docs/grafana-dashboard.json to get the pre-built dashboard. |
| 🆓 | Free & open source | MIT license. The only outbound calls are to your own API keys. |
git clone https://github.com/parmeet10/solveWatchAi.git
cd solveWatchAi
./start.sh --setupThis installs Homebrew, Node.js, Python, Ollama, and all dependencies — then starts the app.
Open the settings page and paste in at least one key:
http://localhost:4000/settings
| Provider | Get a key |
|---|---|
| Groq (fastest — recommended first) | console.groq.com/keys |
| OpenAI | platform.openai.com/api-keys |
| Gemini | aistudio.google.com/app/apikey |
| Claude | console.anthropic.com/settings/api-keys |
| Deepgram (optional — cloud STT) | console.deepgram.com |
Keys are saved to .env in the project root — hot-reloaded, no restart needed.
./start.sh| Shortcut | Action |
|---|---|
⌘ Shift H |
Toggle HUD on/off |
⌘ Shift X |
Toggle listening on/off |
Press Ctrl+C to stop all services.
./start.sh # normal start
./start.sh --setup # install deps + start
./start.sh --setup-only # install deps only
./start.sh --newlogs # clear logs then start
./start.sh --debug # verbose Python transcriber output
./start.sh --newlogs --debug # flags can be combinedThree services run together, managed by start.sh:
Microphone
│
▼
┌─────────────────────────────────────────┐
│ Python Transcriber │
│ │
│ Local mode (default): │
│ VAD → rolling buffer │ Whisper MLX (Apple Silicon) or
│ LocalAgreement-2 streaming decoder │ openai-whisper (Windows/CPU)
│ → stt_partial every 300ms │ Fully offline, no API key
│ → stt_final on 700ms silence │
│ │
│ Deepgram mode (DEEPGRAM_ENABLED=true): │
│ DeepgramListener streams audio │ nova-2 cloud STT, ~300ms latency
│ → stt_partial (live words) │ Built-in speaker diarization
│ → stt_final on utterance end │
└──────────────────┬──────────────────────┘
│ Socket.IO
▼
┌─────────────────────────────┐
│ Node.js Backend │ Express + Socket.IO
│ Assembles prompt │ OpenAI → Groq → Gemini → Claude → Ollama
│ Streams answer tokens ─────┼──► Electron HUD
│ Session memory │
└─────────────────────────────┘
│
▼
┌─────────────────────────────┐
│ Electron HUD │ Frameless, always-on-top overlay
│ 380×460px │ setContentProtection(true)
│ Invisible in screenshare │ Renders tokens as they stream in
└─────────────────────────────┘
Screenshot flow runs in parallel: uploads/ folder → Sharp preprocessing → Tesseract OCR → same AI pipeline → HUD.
setContentProtection(true) is an OS-level API — the same one used by banking apps and DRM video players.
- macOS: maps to
NSWindow.sharingType = NSWindowSharingNone - Windows: maps to
SetWindowDisplayAffinity(WDA_EXCLUDEFROMCAPTURE)
These flags tell the OS compositor to exclude the window from all capture streams — window capture and full-screen capture alike. The capture tool receives a frame that was never composed with the overlay. Your interviewer sees only your desktop.
Confirmed invisible on: Zoom · Google Meet · Microsoft Teams · Loom · OBS Studio · macOS Screenshot · Windows Snipping Tool · Discord Go Live
| SolveWatch | Cluely | Parakeet | |
|---|---|---|---|
| Price | Free (MIT) | $29–49/mo | $20–40/mo |
| API cost | Your keys only | Included (their cloud) | Included (their cloud) |
| Open source | ✅ | ❌ | ❌ |
| Offline STT | ✅ | ❌ | ❌ |
| Custom AI provider | ✅ | ❌ | ❌ |
| Response latency | ~200–400 ms | ~600–1200 ms | ~500–900 ms |
| macOS | ✅ | ✅ | ✅ |
| Windows | ✅ | ✅ | — |
HUD doesn't appear
Press ⌘ Shift H. Check the terminal for Electron startup errors.
No transcription / mic not working
Go to System Settings → Privacy → Microphone and allow Terminal (macOS). Try switching to the small Whisper model.
AI not responding Open settings, check at least one provider is enabled, and use the Test button to verify the key.
Screenshot analysis not working
Set the screenshots folder in settings and confirm it matches where macOS saves screenshots (System Settings → Screenshots → Save to).
Transcriber not starting
Run ./start.sh --setup to recreate the Python venv.
Ollama missing
Run ./start.sh --setup — it pulls llama3.2:1b automatically.
All config lives in .env at the project root — hot-reloaded by the settings page, no restart needed. Copy from .env.example on first setup.
# AI providers (at least one required)
OPENAI_API_KEY=sk-...
GROQ_API_KEY=gsk_...
GEMINI_API_KEY=AIza...
ANTHROPIC_API_KEY=sk-ant-...
# Provider order and which are active
PROVIDER_ORDER=openai,grok,gemini,claude
PROVIDER_ENABLED=openai,grok,gemini,claude
# Models (optional — defaults shown)
MODEL_OPENAI=gpt-4o-mini
MODEL_GROK=llama-3.3-70b-versatile
MODEL_GEMINI=gemini-2.5-flash
MODEL_CLAUDE=claude-sonnet-4-5
OLLAMA_MODEL=llama3.2:1b
# STT — local Whisper model size (tiny/base/small/medium)
STT_MODEL=small
# Deepgram Cloud STT (optional — set to true to use instead of local Whisper)
DEEPGRAM_ENABLED=false
DEEPGRAM_API_KEY=
DEEPGRAM_MODEL=nova-2
# Speaker identification (local Whisper mode only)
SPEAKER_ID_ENABLED=false
SPEAKER_ID_THRESHOLD=0.70
# Observability — Grafana Cloud OTLP (optional)
TELEMETRY_ENABLED=false
OTLP_ENDPOINT=https://otlp-gateway-prod-us-east-0.grafana.net/otlp
GRAFANA_INSTANCE_ID=
GRAFANA_ACCESS_TOKEN=
PORT=4000Existing installs with config/api-keys.json are migrated automatically on the first ./start.sh run.
Planned features — contributions welcome:
- Deepgram Cloud STT (nova-2, ~300ms latency)
- Ollama local LLM fallback
- Grafana Cloud observability (OTel metrics + logs)
- Speaker identification (local Whisper mode)
- Linux support
- Browser extension mode (no Electron required)
- Answer history panel with copy-to-clipboard
- Custom hotkey configuration in settings UI
- Automated release builds (DMG for macOS, EXE installer for Windows)
MIT — free for personal and commercial use.