NovA AI is a generative AI interview coach that helps Latin American tech professionals prepare for Cloud Engineering, DevOps, and ML interviews — in English or Spanish, with real-time feedback and personal progress tracking.
Tech professionals in LATAM face a real gap when preparing for technical interviews at global companies:
- No contextualized practice — most resources are in English and oriented toward US/EU markets.
- No immediate feedback — practicing alone doesn't tell you which concepts you missed or what to study next.
- No progress tracking — there's no way to know if you're actually improving session over session.
NovA AI solves all three using AWS generative AI, fully accessible from the browser with no setup required.
| Feature | Description |
|---|---|
| Conversational AI interviewer | Nova greets you by time of day and conducts the interview like a real conversation, not a form |
| Adaptive questions | Bedrock Nova 2 Lite generates questions and adjusts difficulty in real time based on your performance |
| Instant evaluation | Every answer gets a 0–100 score, strengths, improvement areas, and recommended topics to study |
| 3 technical roles | Cloud Engineer · DevOps Engineer · ML Engineer at Junior and Mid-level |
| Fully bilingual | English and Spanish with fully localized prompts, feedback, and UI |
| Bidirectional voice | Amazon Nova Sonic processes audio in real time over WebSocket — speak and listen without typing |
| Personal dashboard | Session history, average score, score progression chart, and per-area breakdown — persisted per user via DynamoDB |
| PDF report | A downloadable PDF with the full interview breakdown is generated at the end of each session |
| Google Sign-In | Auth via Cognito PKCE + Google OAuth — entirely optional, demo mode works without an account |
Browser (index.html · vanilla JS)
├── Landing → Google Sign-In (Cognito PKCE) / Try Demo
├── Setup → Role · Level · Language
├── Interview → Chat + Voice (WebSocket)
├── Results → Score ring · Area bars · PDF download
└── Dashboard → History table · Sparkline · Area breakdown
│
│ HTTP + WebSocket (persistent process — App Runner)
▼
FastAPI (app.py)
├── /auth/* → Cognito JWT validation · PKCE exchange · logout
├── /api/session/* → Start · Next question · Submit response · End
├── /api/user/*/sessions → Per-user DynamoDB session history
├── /api/report/{id} → PDF download
└── /ws/{id} → Nova Sonic bidirectional audio stream
│
▼
mock_interview_coach/
├── question_generator/ Bedrock Nova 2 Lite — adaptive question generation
├── evaluator/ 0–100 scoring · structured feedback
├── difficulty_adjuster/ Dynamic difficulty per session
├── report_generator/ PDF export via ReportLab
├── voice_interface/ Nova Sonic client · WebSocket handler · audio utils
└── auth/ Cognito JWT · PKCE · DynamoDB helpers
│
▼
AWS
├── Bedrock Nova 2 Lite → Question generation + answer evaluation
├── Bedrock Nova Sonic → Real-time bidirectional voice streaming
├── Cognito Hosted UI → Google OAuth (optional)
├── DynamoDB → nova_users + nova_sessions (silent no-ops when unconfigured)
└── App Runner → HTTPS · auto-scaling · persistent WebSocket
Why App Runner and not Lambda: WebSockets require a persistent process. Lambda has a 29-second max invocation and no support for long-lived bidirectional connections. App Runner scales to zero when idle and serves WebSocket natively — zero ops overhead.
# 1. Clone and install
git clone https://github.com/AndresRJ18/nova-ai-interview.git
cd nova-ai-interview
python -m venv .venv && source .venv/Scripts/activate # Windows
pip install -r requirements.txt
# 2. Configure environment
cp .env.example .env
# Edit .env: set AWS_REGION + AWS_ACCESS_KEY_ID + AWS_SECRET_ACCESS_KEY
# 3. Run (demo mode — no AWS credentials required)
ENABLE_DEV_MODE=true python app.py
# Open http://localhost:8000To run with real AWS (Bedrock-generated questions):
# Without ENABLE_DEV_MODE — uses your AWS credentials
python app.py| Service | Role in the project | Why this and not something else |
|---|---|---|
| Bedrock Nova 2 Lite | Question generation + answer scoring | Low latency, competitive token cost, native support in us-east-1 |
| Bedrock Nova Sonic | Real-time bidirectional voice | The only AWS multimodal model with bidirectional audio streaming over WebSocket |
| App Runner | FastAPI container hosting | Native WebSocket support, scales to zero, automatic HTTPS, no infra to manage |
| Cognito Hosted UI | Google OAuth + PKCE | Removes secret handling from the frontend; the hosted UI handles the OAuth redirect directly |
| DynamoDB | User profiles + session history | PAY_PER_REQUEST, serverless, no capacity planning — perfect for variable hackathon traffic |
| ECR | Container registry | Native integration with App Runner for auto-deploy on every image push |
| Variable | Required | Description |
|---|---|---|
AWS_REGION |
Yes | e.g. us-east-1 |
NOVA_SONIC_MODEL_ID |
Yes | amazon.nova-sonic-v1:0 |
COGNITO_USER_POOL_ID |
No | Enables Google login |
COGNITO_CLIENT_ID |
No | Cognito app client |
COGNITO_DOMAIN |
No | Hosted UI domain |
COGNITO_REDIRECT_URI |
No | OAuth callback URL |
DYNAMO_USERS_TABLE |
No | Default: nova_users |
DYNAMO_SESSIONS_TABLE |
No | Default: nova_sessions |
ENABLE_DEV_MODE |
No | true for local dev without AWS credentials |
On App Runner, credentials are injected via IAM Role — no AWS_ACCESS_KEY_ID needed in production.
ACCOUNT=$(aws sts get-caller-identity --query Account --output text)
REGION=us-east-1
# Create registry and push image
aws ecr create-repository --repository-name nova-interview --region $REGION
aws ecr get-login-password --region $REGION | \
docker login --username AWS --password-stdin $ACCOUNT.dkr.ecr.$REGION.amazonaws.com
docker build -t nova-interview .
docker tag nova-interview:latest $ACCOUNT.dkr.ecr.$REGION.amazonaws.com/nova-interview:latest
docker push $ACCOUNT.dkr.ecr.$REGION.amazonaws.com/nova-interview:latestRequired IAM Role (nova-apprunner-role):
bedrock:InvokeModel,bedrock:InvokeModelWithResponseStreamon*dynamodb:PutItem,dynamodb:GetItem,dynamodb:UpdateItem,dynamodb:Queryon both tables
1. Frontend calls GET /auth/config → receives client_id, domain, redirect_uri
2. PKCE: browser generates code_verifier + code_challenge (stored in sessionStorage)
3. Redirect to Cognito Hosted UI → Google login → back to app with ?code=
4. Frontend POST /auth/callback → {code, redirect_uri, code_verifier}
5. Backend exchanges code for tokens, validates RS256 JWT (JWKS cached), upserts user in DynamoDB
6. Token stored in localStorage → sent as Authorization: Bearer <token> on each request
When Cognito env vars are absent: the Google button stays disabled and demo mode works normally. No errors, no broken screens.
| Area | Improvement |
|---|---|
| Voice in cloud | Replace App Runner with ECS Fargate + ALB — App Runner's Envoy proxy blocks WebSocket upgrades (HTTP 403), preventing Nova Sonic bidirectional streaming in production. ECS + ALB natively supports WebSocket. |
| Voice in cloud (alt) | Add API Gateway WebSocket API in front of the voice endpoint, with a Lambda proxy that maintains the Nova Sonic session. |
| Auth | Add refresh token rotation — current implementation stores the id_token in localStorage without expiry handling. |
| Dashboard | Session history table with score progression chart and per-area breakdown from DynamoDB. Endpoint /api/user/{id}/sessions is designed but UI pending. |
| PDF report | Add charts (score ring, area bars) to the PDF export using ReportLab's drawing API. Currently text-only. |
| Question difficulty | The DifficultyAdjuster is wired but the adjustment signal is not fed back into question generation. Close the loop for truly adaptive interviews. |
| Prompt language | Nova Sonic system prompts still reference "the candidate" in some edge cases. Full second-person audit needed. |
| Tests | Increase test coverage for session_manager and evaluator with integration tests against a local DynamoDB mock. |
Andrés Rodas — Computer Engineering, UPCH · Cloud Computing & AI
Made with love by Andres · Lima, Peru · 2026