Runs on your machine. Nothing is uploaded.
Open Interview builds an interview rubric from the job description you are hiring against, runs a spoken interview using a local model, and refuses to record anything. If the model cannot find a phrase it claims the candidate said, the score is thrown away rather than shown to you.
857
automated tests passing
42/42
IPC request channels guarded
3
local services, loopback only
0
cloud calls by default
v0.1.7 verification: 97 test files, 857 tests, 42 IPC request channels and 3 event channels validated against one shared JSON Schema. The default brain is local; a cloud brain is an explicit per-interview choice.
The problem
Communication, problem solving, teamwork. Competencies that would score a frontend hire and a site reliability engineer identically, because nobody ever looked at the posting.
The local model reads the description, proposes competencies, and keeps the exact sentence each one came from. Anything whose quote is not found in the text is dropped before you ever see it.
A PDF, Word document, Markdown or plain text. It is read on the device and split into labelled sections with real character offsets.
Every bundled interview profile is scored against the text, with the evidence shown. It never picks silently — you confirm the archetype and the seniority band.
Competencies arrive beside the sentence they came from. Profile-seeded ones are marked inferred so you can tell them apart. Nothing interviews a candidate until a person approves it.
Speak naturally. A pause of half a second, one second or two sends the answer to local speech recognition. The model chooses the next question; it never decides when the interview ends.
Weighted per competency, against that competency's own anchors, each score carrying the transcript phrases it rests on.
Scoring
Every competency score must cite one to three exact phrases from what the candidate actually said. The app searches the transcript for each one. A citation that is not found is rejected, the model is asked again once, and if it invents evidence a second time that competency is reported as not scored — named, with the reason, instead of quietly guessed at.
Backend engineer · senior
Reliability & correctness
Named the failure mode before reaching for a fix.
“I made the write path idempotent first, so the retry storm was harmless.”
Operability
Designed for the person on call, not the author.
“If I cannot tell from a dashboard whether the queue is stuck, the dashboard is part of the bug.”
Distributed systems depth
The model cited “led a team of fifty engineers”. That phrase does not appear anywhere in the transcript, so the score was discarded and the competency is reported as not scored.
“led a team of fifty engineers” not found in transcript — discarded
Scores are decision support, not a decision. The weighted overall covers only the competencies that were both asked and scorable.
Privacy
Not a promise in a policy page — the default architecture. The model, speech recognition and voice run on loopback addresses on this device. No cloud endpoint is used unless an administrator explicitly enables an external provider, and the local path needs no key.
The microphone is live only while you are answering. Samples are converted in memory and handed straight to the local recogniser.
The core path runs without an account. A cloud provider key can be stored in the OS keychain only when an administrator enables an external provider; providers are optional.
Every local endpoint is checked to be loopback, and redirects are refused, so a misconfigured local address cannot quietly forward an interview. External providers are opt-in and redacted before a payload leaves the device.
Transcripts and reports are saved beside the rubric library on this machine, listed on the home screen, and deleted on request — including a single confirmed delete-everything.
Providers
Run the whole interview on loopback with Ollama, whisper.cpp and Piper. Or choose a configured provider as the interview brain on Home; it handles the opening, replies and scoring, with the local stack as the fallback.
Local
The default path. Language, speech recognition and voice all run on this device, with no account and no cloud endpoint.
External · LLM
Use a GPT model for interview questions. The key stays in the OS keychain; local Ollama remains the fallback.
External · LLM
One key for many hosted models. Choose a model id and the router will fail over to the local model if the provider is down.
External · LLM
OpenCode's OpenAI-compatible gateway, configured with its base URL and model id.
External · Speech + LLM
Sarvam chat plus Saaras speech recognition and Bulbul speech output for Indian languages.
External · Gnani voice
Gnani Prisma speech recognition and Timbre speech output, wired into the live interview with a Gnani profile and key.
Add a profile in Settings → Providers and fallbacks. The key is checked before saving; OpenAI-compatible providers load their model list, and Gnani reports whether Prisma and Timbre are in scope. The workspace policy decides whether external providers may be selected; the local stack is always the fallback. Keys never leave the OS keychain.
Built with Gnani
This build was made for the Great Indian AI Internship Challenge. When a Gnani profile is enabled, the live interview uses Gnani's voice models for both sides of the conversation, with the local Whisper and Piper path as the fallback.
Gnani · Speech to text
Candidate audio is transcribed with
POST /stt/v3, model prisma-v2.5, and
the X-API-Key-ID header. Prisma v3.0 is available
for indic-auto and mixed-language audio.
Gnani · Text to speech
The interviewer's question is voiced with
POST /api/v1/tts/inference, model
timbre-v2.5, voice Nalini, language
en-IN. Timbre is currently in beta.
Fallback
If Gnani is unavailable or the workspace is local-only, the router falls back to local speech recognition and voice on the next turn. The interview does not stop.
Settings → Providers and fallbacks → Add provider → Gnani (Prisma + Timbre). Get a key at app.gnani.ai and read the Gnani API docs.
Questions
No, not by default. The language model, speech recognition and voice run on loopback addresses on your machine. An external provider is only used when an administrator enables it and the interviewer chooses it for that interview; the choice and its data boundary are shown before the session starts.
Yes. The local path uses Ollama, whisper.cpp and Piper, and needs no account or key. You need the models installed once; after that the interview itself does not call a cloud endpoint.
Every competency score must cite one to three exact phrases from the transcript. The app searches the transcript for each phrase. A citation that is not found is rejected, the model is asked again once, and if it invents evidence a second time the competency is reported as not scored rather than guessed.
OpenAI-compatible providers such as OpenAI, OpenRouter, OpenCode and Sarvam chat, plus Gnani Prisma speech recognition and Timbre speech output. A key is checked before it is saved; OpenAI-compatible providers load their model list and Gnani reports whether Prisma and Timbre are in scope. Keys are stored in the OS keychain, and local Ollama remains the fallback.
Windows 10 or 11 on x64, Ollama with a model installed, and whisper.cpp and Piper for local speech. The installer is not code-signed yet, so Windows SmartScreen may warn; verify the published SHA-256 before running it.
On your machine, beside the rubric library. Audio is converted in memory and is not written to disk. Transcripts and reports can be deleted from the app, including a single confirmed delete-everything action.
Not yet. The current release is Windows x64 only: an NSIS installer, a portable ZIP and an enterprise MSI.
Bring your own model. The app shows you which local models are installed and tells you exactly which service failed if one does.
Windows may warn about this build. It is not code-signed yet. If SmartScreen says “Windows protected your PC”, choose More info → Run anyway. If the browser asks, choose Keep. Verify the SHA-256 below before running it.
Other builds and verification
Security checks
Each artifact is scanned before publishing. The SHA-256 beside each file matches the checksum file. VirusTotal reports 0/67 engines on the installer, 0/64 on the portable ZIP and 0/59 on the MSI. The links above open the exact hashes.
VirusTotal may show generic Sigma behavior matches on the NSIS
installer because it extracts its own plugin DLLs to
%TEMP% and checks for a running copy with PowerShell.
That is standard electron-builder installer behavior, not malware.
If your endpoint blocks the NSIS installer, use the MSI or the
portable ZIP.
Version 0.1.7 for Windows on x64. Requires Ollama, whisper.cpp and Piper running locally. Full setup steps ship with the app.