AI-powered job application tracker that learns from your rejections.
Live demo: https://hardikjp7.github.io/applytic
Applytic tracks every job application you submit, detects patterns across rejections (which resume version converts best, which source channel works, which company sizes respond), and uses Amazon Bedrock to turn that data into actionable coaching — delivered as a chat interface and a weekly email digest.
Built end-to-end on AWS as a production-grade portfolio project. Every service is serverless, infrastructure is code, and every push auto-deploys via GitHub Actions.
I was job hunting and had no data on why I was getting rejected. I had a spreadsheet with company names and "rejected" written next to most of them — but no signal on why. Was it my resume? The channel? The company size?
So I instrumented my own job search. Every application became a data point. After a few weeks I had enough data to see that my v1-generic resume had a 0% response rate from enterprise companies, while v3-ml-focused was getting interviews from startups via referrals. That's the kind of insight you can act on.
| Layer | Service |
|---|---|
| Frontend | React 18, TypeScript, Vite, Tailwind CSS, react-markdown |
| Auth | Amazon Cognito (email + JWT) |
| API | API Gateway REST + Lambda (Python 3.12, ARM64) |
| AI / ML | Amazon Bedrock — Amazon Nova Lite |
| Database | DynamoDB — single-table design, PAY_PER_REQUEST |
| Storage | S3 — resume versioning + frontend hosting |
| CDN | CloudFront (450+ edge locations) |
| Scheduling | EventBridge cron (Monday 8am UTC) |
| Amazon SES | |
| IaC | AWS CDK v2 TypeScript |
| CI/CD | GitHub Actions |
Application tracking
- Log applications with company, role, source channel, resume version, company size, job description URL
- Kanban board with drag-and-drop status updates (Applied → Screened → Interview → Offer / Rejected)
- Click any card to view full detail, edit all fields, change status, see timeline
- Search by company/role, filter by source channel
- Color-coded left border per status on kanban cards for instant visual scanning
AI insight engine
- Pattern analysis across 6 dimensions: source channel, company size, resume version, role seniority, weekly velocity, status funnel
- Response rate computed per bucket — shows exactly which resume version or source is working
- AI coaching chat powered by Bedrock — answers questions like "why am I getting ghosted?" using your actual data as context, not generic advice
- Markdown rendering in chat responses — bold, lists, and structure render properly
- Weekly email digest every Monday with stats + one AI-generated personalised tip
Resume version tracker
- Upload multiple PDF versions to S3 via presigned URLs (never passes through Lambda)
- Tag each application with which version was used
- Analytics shows conversion rate per version side-by-side
UI / UX
- Full dark mode with system preference detection, persisted to localStorage
- Mobile responsive — hamburger sidebar on small screens, responsive grids throughout
- Loading skeletons on every page instead of blank states
- Meaningful empty states with calls to action
- Toast notifications (top-center)
applytic/
├── cdk/ # AWS CDK v2 stack — all infrastructure as code
│ ├── bin/app.ts
│ └── lib/applytic-stack.ts
├── lambdas/ # Python 3.12 Lambda handlers
│ ├── applications/ # CRUD + presigned S3 URL generation
│ ├── insights/ # Pattern analysis + Bedrock AI coaching
│ └── digest/ # Weekly SES email digest
├── frontend/ # React + Vite + Tailwind CSS
│ └── src/
│ ├── components/ # Kanban, Analytics, Chat, Sidebar, Resume
│ ├── hooks/ # useApplications (data fetching + state)
│ ├── lib/ # API client, Amplify config, theme util
│ ├── pages/ # Dashboard
│ └── types/ # Shared TypeScript types
├── tests/ # 48 pytest unit tests
│ ├── conftest.py
│ ├── test_applications.py
│ └── test_insights.py
├── scripts/
│ └── seed_data.py # Populates DynamoDB with 20 realistic demo apps
└── .github/workflows/
└── deploy.yml # CI/CD: test → deploy backend → deploy frontend
Table name: applytic
| Entity | PK | SK | GSI1PK | GSI1SK |
|---|---|---|---|---|
| Application | USER#userId |
APP#appId |
USER#userId |
DATE#timestamp |
| Status event | APP#appId |
EVENT#timestamp |
— | — |
Access patterns:
- List all applications for user → GSI1 query on
USER#userIdsorted by date - Get single application → Main table get on
USER#userId+APP#appId - Get status history for an app → Query
APP#appIdwithEVENT#prefix
Every status change writes a STATUS_EVENT record, creating a full audit trail of the application journey.
Single-table DynamoDB — two entity types (APPLICATION, STATUS_EVENT) in one table with composite keys. Every query is O(1). No joins. Scales to millions of requests/sec with no configuration changes. This demonstrates understanding of NoSQL access pattern design rather than SQL-in-NoSQL thinking.
Pattern analysis before LLM — the insights Lambda computes structured metrics (response rates per bucket) before calling Bedrock. The LLM receives hard numbers as context, not raw records. This makes advice data-driven and specific rather than generic — it's the difference between "apply to more startups" backed by actual numbers vs. a guess.
Serverless-first — no EC2, no containers, no idle servers. Lambda + API Gateway + DynamoDB means ~$0 at low volume and automatic scaling. EventBridge replaces a cron server entirely. The entire system costs roughly $2-5/month at 100 active users.
ARM64 Lambda — all functions run on Graviton2 (ARM64) for ~20% cost reduction and faster cold starts vs x86.
Presigned S3 URLs for resume upload — the Lambda generates a presigned URL and returns it to the client. The file then uploads directly from the browser to S3 — never passes through Lambda. This avoids Lambda payload limits and keeps upload fast.
Multi-model Bedrock support — the insights Lambda detects whether the configured model is Anthropic or Amazon (based on model ID prefix) and adjusts the request/response format accordingly. This means switching models requires only an environment variable change.
| Metric | Value |
|---|---|
| API throughput | 10,000 req/sec (API Gateway default) |
| DynamoDB SLA | 99.999% availability, millions req/sec |
| Lambda cold start | ~300-400ms (Python 3.12 ARM64) |
| CloudFront edge locations | 450+ worldwide |
| Cost at 0 users | ~$0/month |
| Cost at 100 users | ~$2-5/month |
| Cost at 1,000 users | ~$15-30/month |
| Test suite runtime | under 2 seconds (48 tests) |
| Full CDK deploy from scratch | under 3 minutes |
Every push to main triggers three jobs in parallel after tests pass:
push to main
└── test (pytest, 48 tests)
├── deploy-backend (cdk deploy)
└── deploy-frontend (npm build → s3 sync → cloudfront invalidation)
Pull requests only run tests — no deploy. This means every PR is validated before merge.
- Node.js 18+
- Python 3.12
- AWS CLI configured (
aws configure) - AWS CDK CLI:
npm install -g aws-cdk
cd cdk
npm install
cdk bootstrap
cdk deploySave the outputs — you need ApiUrl, UserPoolId, UserPoolClientId for the frontend .env.
cd frontend
cp .env.example .env
# fill in .env with CDK outputs
npm install
npm run devOpens at http://localhost:5173
cd frontend
npm run build
aws s3 sync dist/ s3://applytic-frontend-YOUR_ACCOUNT_ID --delete
aws cloudfront create-invalidation --distribution-id YOUR_DIST_ID --paths "/*"cd scripts
pip install boto3
python seed_data.py --user-id YOUR_COGNITO_SUBTo get your Cognito sub: sign in → DevTools → Application → Local Storage → find idToken → paste at jwt.io → copy sub field.
pip install pytest boto3
python -m pytest tests/ -v| Variable | Description |
|---|---|
VITE_API_URL |
API Gateway base URL from CDK output |
VITE_USER_POOL_ID |
Cognito User Pool ID |
VITE_USER_POOL_CLIENT_ID |
Cognito App Client ID |
VITE_AWS_REGION |
AWS region (us-east-1) |
Never commit .env — it is gitignored. GitHub Actions injects these values via repository secrets at build time.
This section documents every real issue hit during the build — useful for anyone replicating this project and for interviews where "tell me about a hard bug" comes up.
1. TypeScript error on CDK deploy — Partial<FunctionProps> type
Using Partial<lambda.FunctionProps> to share Lambda defaults made runtime optional, but the Function constructor requires it as non-optional. TypeScript rejected the spread.
Fix: removed the shared defaults object entirely and added runtime explicitly on each Lambda definition.
2. esbuild crash on Windows (npm run dev)
esbuild's native binary gets corrupted on Windows when Node versions mismatch or node_modules gets into a bad state.
Fix: delete node_modules and package-lock.json, reinstall with npm install, use Node 18 or 20 LTS only.
3. pytest module name collision
Both Lambda handlers are named handler.py. Python caches the first import under the name handler, so the second test file imports the wrong one.
Fix: use importlib.util.spec_from_file_location to load each handler by absolute file path, registering each under a unique module name (applications_handler, insights_handler).
4. Timezone-aware datetime bug in velocity calculation
dateApplied is stored as YYYY-MM-DD (date string). The velocity calculation tried to subtract it from a timezone-aware datetime.now(timezone.utc), causing TypeError: can't subtract offset-naive and offset-aware datetimes.
Fix: detect date-only strings (length 10) and parse with .replace(tzinfo=timezone.utc) before subtraction.
5. Chat history lost on page navigation
React unmounts the component on route change, so useState inside CoachChat reset to the initial message every time the user navigated away.
Fix: lifted the messages state to App.tsx and passed it as props to CoachChat. State now lives at the router level and survives navigation.
6. GitHub Actions npm cache error
cache: 'npm' in setup-node requires a package-lock.json to exist. The lockfile wasn't committed.
Fix: removed cache and cache-dependency-path from both Node setup steps, changed npm ci to npm install.
7. import.meta.env TypeScript error in CI
The build failed in CI with Property 'env' does not exist on type 'ImportMeta'. Vite provides this type via vite/client but the tsconfig didn't include it.
Fix: added "types": ["vite/client"] to compilerOptions in tsconfig.json.
8. Bedrock model end-of-life (anthropic.claude-3-5-sonnet-20241022-v2:0)
The originally configured model reached AWS end-of-life mid-project. Lambda started throwing ResourceNotFoundException.
Fix: updated BEDROCK_MODEL_ID environment variable. Learned to use aws logs tail for Lambda debugging and aws lambda update-function-configuration for immediate hotfixes without CDK redeploy.
9. Bedrock inference profile requirement
Newer Claude models (3.7 Sonnet) require an inference profile ID (us.anthropic.claude-3-7-sonnet-20250219-v1:0) rather than the direct model ID. On-demand throughput isn't supported with the bare model ID.
Fix: prefix with us. for cross-region inference profiles.
10. IAM policy not covering inference profile ARN
The CDK IAM policy granted access to arn:aws:bedrock:region::foundation-model/anthropic.* but inference profiles have a different ARN format (arn:aws:bedrock:region:accountId:inference-profile/*). The Lambda got AccessDeniedException.
Fix: changed the Bedrock IAM policy resource to * to cover both foundation models and inference profiles.
11. AWS Marketplace subscription required for newer models
Both Claude 3.7 Sonnet and Claude Haiku 4.5 required AWS Marketplace subscription acceptance before use. This is a newer AWS requirement for certain Anthropic models.
Fix: switched to Amazon Nova Lite (amazon.nova-lite-v1:0) which works without Marketplace subscription on all AWS accounts.
12. Amazon Nova uses different request/response format
Nova uses a completely different API schema from Anthropic. Sending Anthropic-format requests (anthropic_version, max_tokens at top level) to Nova caused ValidationException: extraneous key [max_tokens] is not permitted.
Fix: updated chat_with_coach() to detect the model family from the model ID and use the appropriate request/response format for each. The Lambda now supports both Anthropic and Amazon model families transparently.
13. Pie chart labels clipped at top of card Recharts renders percentage labels outside the pie slices by default. The top label overflowed the card boundary and got clipped by CSS overflow. Fix: replaced external labels with a custom render function that draws white percentage text inside each slice at the midpoint radius. Labels that would be too small (< 5% slice) are skipped automatically.
Lambda API Gateway DynamoDB S3 CloudFront Cognito Bedrock SES EventBridge CloudWatch IAM CDK
Contributions are welcome! See CONTRIBUTING.md for setup instructions, branching conventions, and the PR process.
Hardik — Github