Newshot

Show HN: Interviewer assistant – AI open source

Original text(via HN AI)

AI Interview Assist

题图
题图

An AI assistant for interviewers. A bot joins the interview video call (Microsoft Teams for now soon to come Zoom or Google Meet), listens, and privately coaches the interviewer in a web app: a live transcript, a score for every answer (correctness, depth, clarity, relevance, with the weak phrases quoted and a better answer suggested), and follow-up questions to ask next. The assistant never speaks in the meeting, and only the interviewer sees the coaching in web app.

Meeting access and live transcription come from Recall.ai. AI analysis uses Anthropic Claude and Azure OpenAI models. This repo contains the backend that schedules the bot, receives the transcript and runs the AI, plus the web app interviewers use.

Consent matters. A visible bot joins and the meeting is recorded and transcribed. Whoever runs this must tell every participant and get their permission first, and comply with the recording and privacy laws that apply to them. Data is processed by Recall.ai and your AI providers, and stored in your own database. You run it at your own risk; see the in-app Terms and Privacy pages (web/src/pages/Legal.tsx) and adapt them to your situation.

Read this first: risks and how you can help

  • Vibe-coded, open source, use at your own risk. This application was built largely with AI assistance. I took care to close gaps (there are tests, and sign-in, access control and webhook verification were reviewed), but it has not had an independent security audit and there are surely bugs. Do not treat its scores or suggestions as the basis for a hiring decision, and do not rely on it to keep or recover data (see the Terms in the app).
  • Get consent. Tell candidates and everyone on the call that a bot will join and that the meeting is recorded, transcribed and analysed by AI, and get their permission first.
  • It is self-hosted and I pay for it. The hosted instance runs on my own infrastructure: I pay for the meeting bot (Recall.ai) and the AI models. If it helps your interviews, please consider contributing: Pay what you like through Stripe ($5, $7 or $10) from the app. Nothing is gated on it. The app shows a small, optional "found it useful?" line under the interview setup form and a card on the Interviews page; the $ amounts appear only when STRIPE_SECRET_KEY is set. Try it on your own interviews and report bugs or rough edges as GitHub issues. Contribute code, tests, docs or fixes by pull request, or help maintain it. Run your own copy (below) with your own Recall.ai, Infra & AI keys.
  • Pay what you like through Stripe ($5, $7 or $10) from the app. Nothing is gated on it. The app shows a small, optional "found it useful?" line under the interview setup form and a card on the Interviews page; the $ amounts appear only when STRIPE_SECRET_KEY is set.
  • Try it on your own interviews and report bugs or rough edges as GitHub issues.
  • Contribute code, tests, docs or fixes by pull request, or help maintain it.
  • Run your own copy (below) with your own Recall.ai, Infra & AI keys.

Docs: plan · hosting & DNS · inbound e-mail · Recall integration · data & backups · end-to-end flow · why not our own Teams bot.

What you need

NeedWhyRequired?
Node.js 24 and npm (what the Docker image uses)Run the backend and web appYes
Recall.ai account and API keyBot that joins meetings and streams the transcriptFor real meetings (not for local dev)
Anthropic API key and/or Azure OpenAI resourceThe AI that scores answers and suggests questionsFor coaching
Microsoft Entra app registration"Sign in with Microsoft"For real sign-in (not for local dev)
A public HTTPS URL for the backendRecall and SendGrid deliver webhooks to it (a tunnel such as ngrok works for testing)For real meetings
SendGrid account + a subdomainOptional: lets people invite the bot by e-mail ( iv-<alias>@your-subdomain )Optional
Stripe accountOptional "support us" paymentOptional
A Linux VM with DockerProduction hosting (Caddy gives automatic HTTPS)For production

How it works


        interviewer adds [email protected] to the meeting invite
        │  (no mailbox: the subdomain's MX record points at SendGrid)
SendGrid Inbound Parse ──POST──►  backend /inbound/sendgrid/<secret>
                      │  parse link (Teams/Zoom/Meet) + calendar data, match alias → paid interview
                      ▼
                 Recall: create bot (scheduled for the start time, or immediate)
                      │
 meeting ◄──joins──  Recall bot ──live transcript + status──►  backend /webhooks/recall  (HMAC-verified)
                                                                   │
                                                                   ▼
                                                    status tracking, transcript segments → coaching (next)
      

Layout

PathWhatState
backend/TypeScript (Fastify). Inbound invite e-mail, Recall client, bot scheduling, webhook receiver, live AI coaching and scoring, SQLite persistenceTested; verified live end to end
web/React + Vite single-page app: Microsoft sign-in, interview setup, live transcript, scores and suggested questionsBuilt
deploy/Docker Compose: backend + Caddy (automatic HTTPS), nightly DB backupUsed in production
infra/sendgrid/Script to create the SendGrid Inbound Parse hostWritten
infra/azure/provision.sh : resource group + Key VaultUsed
teams-app/Optional Teams chat app for private coaching messages (chat only)Manifest packs, not sideloaded
doc/Plan, flow, Recall integration guide, research history

Tag teams-media-bot-spike holds the earlier own-Teams-media-bot attempt (it could join meetings but Microsoft does not allow link-joined bots to record or transcribe).

Quick start (local, no accounts needed)

Local dev mode skips Microsoft sign-in and never creates a real bot, so you can run the whole app on your machine.


        git clone https://github.com/blackarck/intaiassist.git && cd intaiassist

# 1. backend (http://localhost:3000)
cd backend
npm install
cp .env.example .env            # git-ignored; set INTAKE_TOKEN to any long random string
DEV_AUTH=1 RECALL_API_KEY= ALLOWED_ORIGINS=http://localhost:5173 npm run dev

# 2. web app (http://localhost:5173), in a second terminal
cd web
npm install
cp .env.example .env.local
# in .env.local set: VITE_API_URL=http://localhost:3000   [email protected]
npm run dev
      

Open http://localhost:5173. To play a fake meeting without Recall, see "Local development" in web/README.md. Add ANTHROPIC_API_KEY (and/or the AZURE_OPENAI_* pair) to backend/.env to get real AI scoring. DEV_AUTH=1 is refused when NODE_ENV=production, and VITE_DEV_USER must never be set in a deployed build.

Configuration

Backend settings live in backend/.env (see backend/.env.example for the full, commented list). The main ones:

VariablePurpose
INTAKE_TOKENOperator bearer token for admin/dev API calls. Long random string.
MS_CLIENT_IDEntra app (client) id for sign-in. Same value as the web build's VITE_MS_CLIENT_ID .
ALLOWED_ORIGINSWeb app origin(s) allowed to call the API (CORS), e.g. https://app.example.com .
RECALL_API_KEY , RECALL_REGION , RECALL_WEBHOOK_SECRETRecall.ai credentials, region and webhook verification secret.
PUBLIC_BASE_URLPublic https URL of the backend; Recall posts transcripts to <PUBLIC_BASE_URL>/webhooks/recall .
ANTHROPIC_API_KEY , AZURE_OPENAI_ENDPOINT , AZURE_OPENAI_API_KEYAI providers. Model choice per task is in backend/src/ai/ .
ALIAS_DOMAIN , INBOUND_TOKENOptional: SendGrid Inbound Parse for e-mail invites.
ADMIN_SUBJECTSComma-separated user ids who may change AI settings (get yours from GET /api/me after signing in).
DATABASE_PATHSQLite file. Unset means in-memory (lost on restart).
STRIPE_SECRET_KEY , WEB_BASE_URLOptional support payment. Leave blank to turn it off.

Web settings are build-time variables in web/.env.local (see web/.env.example): VITE_API_URL, VITE_MS_CLIENT_ID, VITE_MS_AUTHORITY.

Microsoft sign-in (Entra)

  • In Entra ID create an app registration with account type any organizational directory and personal Microsoft accounts (multi-tenant), so interviewers from any company can sign in.
  • Add platform Single-page application with redirect URIs http://localhost:5173/redirect.html and https://<your-web-domain>/redirect.html.
  • Set the client id as MS_CLIENT_ID (backend) and VITE_MS_CLIENT_ID (web), with VITE_MS_AUTHORITY=https://login.microsoftonline.com/common.

More detail in web/README.md.

Hook up your own Recall.ai and AI models

Running your own copy means using your own accounts. Everything is set in backend/.env (or /opt/intassist/.env in production), then restart the backend (npm start, or docker compose restart backend).

Recall.ai (the meeting bot)

  • Create an account at recall.ai and note your region (us-west-2, us-east-1, eu-central-1 or ap-northeast-1). Keys only work in the region they were created in.
  • In the Recall dashboard create an API key, and under Developers add a webhook endpoint https://<your-backend-domain>/webhooks/recall; copy its verification secret (whsec_...).
  • Set in the env file: RECALL_API_KEY=<your api key> RECALL_REGION=us-west-2 RECALL_WEBHOOK_SECRET=whsec_... PUBLIC_BASE_URL=https://<your-backend-domain> # no trailing slash, must be public HTTPS # optional: BOT_NAME=Interview Assist TRANSCRIPT_LANGUAGE=en Set all of these or none: with RECALL_API_KEY set but the others missing, the backend reports what is missing. With none set, no real bot is created. For local testing use a tunnel (ngrok, Cloudflare Tunnel) as PUBLIC_BASE_URL.

Set in the env file:


        RECALL_API_KEY=<your api key>
RECALL_REGION=us-west-2
RECALL_WEBHOOK_SECRET=whsec_...
PUBLIC_BASE_URL=https://<your-backend-domain>     # no trailing slash, must be public HTTPS
# optional: BOT_NAME=Interview Assist   TRANSCRIPT_LANGUAGE=en
      

Set all of these or none: with RECALL_API_KEY set but the others missing, the backend reports what is missing. With none set, no real bot is created. For local testing use a tunnel (ngrok, Cloudflare Tunnel) as PUBLIC_BASE_URL.

  • Details and troubleshooting: doc/recall-integration.md.

AI models (the scoring and coaching)

The backend calls models through two providers. Set the keys for the ones you use:


        ANTHROPIC_API_KEY=sk-ant-...                       # Claude (console.anthropic.com)
AZURE_OPENAI_ENDPOINT=https://<resource>.openai.azure.com   # optional: Azure OpenAI / Foundry
AZURE_OPENAI_API_KEY=...
      
  • Anthropic only is enough. Several tasks default to an Azure model first and fall back to Claude, so with no Azure settings they still work but each call first tries and fails on Azure. Avoid that by reassigning those tasks (next point).
  • Choose which model does what without redeploying: sign in as an admin and open AI settings. Admins are the users listed in ADMIN_SUBJECTS (sign in once, call GET /api/me, copy subject, put it in ADMIN_SUBJECTS, restart). For each task (live coach, answer scoring, opening questions, scorecard, assist) pick a primary and fallback model.
  • Model list: the available models are in backend/src/ai/models.ts (CATALOG). Azure entries are deployment names on your Azure resource, so edit them to match your deployments. Defaults per task are in backend/src/ai/tasks/*.ts.
  • Restart the backend after changing keys. Check the backend logs if a task fails; a missing key shows as ANTHROPIC_API_KEY is not set or AZURE_OPENAI_ENDPOINT / AZURE_OPENAI_API_KEY are not set.

Deploying to your own environment

Production is one Linux VM running Docker Compose: Caddy (HTTPS, serves the built web app) and the backend (SQLite on a Docker volume). Outline:

  • Create a VM with Docker and point DNS for your API domain and web domain at it.
  • Put the backend settings above into /opt/intassist/.env, plus DOMAIN (API host) and WEB_DOMAIN (web host).
  • Build the web app (cd web && VITE_API_URL=https://<api-domain> VITE_MS_CLIENT_ID=<id> VITE_MS_AUTHORITY=https://login.microsoftonline.com/common npm run build) and copy web/dist to /opt/intassist/web.
  • From deploy/: docker compose --env-file /opt/intassist/.env up -d --build.
  • In Recall, point the webhook at https://<api-domain>/webhooks/recall. Optionally run infra/sendgrid/ for e-mail invites.

The scripts in infra/azure/ automate this on Azure (VM, Key Vault for secrets, deploy); they contain our own resource names, so treat them as a worked example and change the names. Full walk-through in doc/hosting.md. Do not deploy or restart during live interviews: a restart drops in-flight webhooks for a few seconds.

Backend only: try the intake flow

Without RECALL_API_KEY the backend uses a log-only bot controller. With DEV_AUTH=1 you can exercise intake with curl:


        curl -X POST localhost:3000/api/dev/interviews -H "authorization: Bearer $INTAKE_TOKEN" -H 'content-type: application/json' \
  -d '{"alias":"7KQ2","ownerEmail":"[email protected]"}'
curl -X POST localhost:3000/api/intake/link -H "authorization: Bearer $INTAKE_TOKEN" -H 'content-type: application/json' \
  -d '{"alias":"7KQ2","link":"https://meet.google.com/abc-defg-hij"}'
      

Recall can only deliver webhooks to a public HTTPS URL, so a real end-to-end test needs the backend hosted or tunnelled (see doc/recall-integration.md).

Optional support payment (Stripe)

People can optionally pay $5, $7 or $10; nothing is gated on it. To turn it on, put your keys in backend/.env (git-ignored):


        STRIPE_SECRET_KEY=sk_test_...        # https://dashboard.stripe.com/apikeys
WEB_BASE_URL=http://localhost:5173   # where Stripe sends people back
      

Leave STRIPE_SECRET_KEY empty and the web app shows no payment amounts (the small "found it useful?" line stays, without them).

If you run your own copy, the contribution prompts say this instance costs infrastructure, meeting-bot and AI money to run. Reword them for your own situation in web/src/components/Support.tsx (SupportCard and SupportNote). The Stripe keys are yours, so payments go to your Stripe account. To remove the prompts, delete <SupportCard /> from web/src/pages/Interviews.tsx and <SupportNote /> from web/src/components/SetupForm.tsx. The backend creates a Stripe Checkout session (POST /api/me/support/checkout); amounts are fixed in backend/src/support.ts. For the Azure deploy, store the key as the Key Vault secret stripe-secret-key. Payments are not tracked in the database: the Stripe dashboard is the record.

Tests


        cd backend && npm test && npm run typecheck
cd web && npm test && npm run typecheck
      

License and sponsor

Source: https://github.com/blackarck/intaiassist. Released under the Apache License 2.0: free to use, modify and distribute, including commercially. Contributions are welcome, see CONTRIBUTING.md. Interview Assist is sponsored by Proso.ai.

The text above was extracted automatically from the source page and may miss images, tables or code blocks. The original is authoritative.

Published by HN AI
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗