AI Enterprise SolutionsGPT & Claude

RedFlag AI

THE USER Customer-support managers and QA leads at companies that talk to clients through chatbots, phone and email: banks, telecoms, e-commerce and service providers. They own customer satisfaction, but they can't read every conversation. THE PROBLEM Today, support quality is checked by hand. A QA team reviews a small random sample of conversations, days or weeks later, using subjective checklists. Most conversations are never reviewed, so an angry or about-to-leave customer is only noticed after a complaint, a bad review or a cancellation. Managers lack a live, consistent picture of how satisfied clients are across channels, customers and support agents. OUR SOLUTION Client Pulse AI receives every conversation through a simple API (one key per company). AI scores each one from 0 to 100 within seconds and explains why. A dashboard shows trends by date, channel, customer and support assistant, and the admin gets an email when a conversation scores below 20, so the team can step in while the customer can still be saved.

Open the demo Video GitHub

Preliminary score

67.0/ 100Promisingaverage of 2 judges

Value for the user19.0 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI23.0 / 30
A working core scenario, and what AI actually contributes to the result.
Quality testing14.5 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility6.5 / 15
Data requirements, running costs and a clear next step.
Originality4.0 / 10
A meaningful difference from common solutions.

What the judges wrote

Claude Opus 5.5Preliminary63/100
Strengths
You target a specific user, support managers and QA leads, with a real pain: only a small sample of conversations gets reviewed, and days late. Your prototype appears complete end to end. It has a public API with hashed per-company keys, background Claude scoring with written reasons, dashboard filters by channel, customer and assistant, and email alerts below a score of 20, all deployed on Render. Your testing log is unusually concrete and honest. You used a deterministic fake model to force outages and furious customers, and you documented real failures and their fixes: the config bug that wiped data, the 401 vs 400 logout issue, the SMTP host error, and the responsive layout bugs.
To improve
Almost all of your testing checks the software plumbing, not whether the AI's scores are right. Have a person score 50–100 varied conversations, compare the results with Claude's scores, and report agreement, false alarms and missed angry customers. Your comparison with today's process is a qualitative table; adding even a rough measure (time saved per QA reviewer, or how many at-risk customers were flagged earlier) would turn it into an adoption metric. You do not mention running costs per scored conversation with an Opus-class model, or a next step such as a pilot partner or real anonymised transcripts. Covering both would greatly strengthen your feasibility case.
GPT-6 AstraPreliminary71/100
Strengths
Your asynchronous scoring-to-alert workflow addresses a clear support-QA problem without making customers wait for model evaluation. Your test report provides concrete measurements and failure cases, including AI outages, revoked keys, authentication errors, and a 95-record migration. The API, dashboard filters, and verified-email alerts describe a credible end-to-end workflow rather than an isolated AI feature.
To improve
Your deterministic fake validates application behavior, but not whether Claude's satisfaction scores or the below-20 alert threshold are reliable; evaluate these against human-rated conversations and report alert precision and recall. Compare review time and missed urgent cases with manual QA rather than relying only on qualitative comparisons. Add per-1,000-conversation costs, customer-data retention and redaction controls, and a scoped pilot with a measurable adoption target. Clarify your advantage over existing sentiment-monitoring tools, ideally through a validated, domain-specific QA rubric.

The team

Sona Naghdaliyeva
data
Nijat AlizadaLead
developer
Jahangir Alisoy
developer

Disclosure

Required by the rules. Judges verify it against the repository.

Models: Anthropic Claude (claude-opus-5-5) via the official Anthropic Java SDK: generates chatbot replies in the simulator and scores each conversation 0–100 with a written reason. Development was assisted by Claude Code (AI coding assistant).
Data: Synthetic only. Demo customers and support assistants are fictional records seeded by the app. Conversations are created live in the built-in simulator. No real customer data is used.
Components: Backend: Spring Boot 4, Spring Security, Spring Data JPA / Hibernate, PostgreSQL, jjwt (JWT), Jakarta Mail, Lombok, springdoc-openapi. Frontend: React 19, TypeScript, Vite, React Router, lucide-react icons. Deployment: Docker, nginx, Render.
Built during the event: Confirmed

Quality testing

HOW WE TESTED - Backend: every endpoint called on a running server with real PostgreSQL. - AI: a deterministic "fake Claude" returns fixed scores (8/58/82) by keyword, so tests are repeatable, free, and can force a furious customer or an AI outage. - Frontend: Playwright on the real app at 360/390/768/1024/1440 px, light and dark. We read every screenshot. - Live: deployed on Render with a real Claude key and real SMTP. 1) Customers never wait for the AI judge The chat replied in 0.47 s while scoring (slowed to 3 s) ran in the background: Pending, then Completed. Broke: with the AI returning HTTP 500, the request was saved as FAILED with the reason, and the chat kept working. Today: QA reviews conversations days later. We score during the conversation. 2) A config bug would have wiped every user The database settings were indented one level too deep, so Spring ignored them: no tables, and data dropped on every restart. Fixed, and switched to safe schema updates. Users now survive restarts. 3) Live data upgraded with zero loss Added Channel, Customer and Support Assistant to a database with 95 scored conversations. All 95 migrated automatically to "Chat", with no manual SQL. 4) Security: we tried to break in Wrong password, unknown user, garbage token, refresh token used as access token and the reverse: all rejected (401). Expired sessions refresh silently. Broke: "wrong current password" in Settings first returned 401, which would have logged the user out. Changed to 400. 5) A deleted API key stops the simulator instantly Deleted the key mid-simulation: the backend refused (403) BEFORE calling the AI, so no cost. The UI showed "Simulation is paused" and kept the typed message. Also: keys stored only as SHA-256 hashes, secret shown once, 6th key refused, another user's key returns 404, fake key on the public API returns 401. 6) Production broke, and told us exactly why - "x-api-key header is required": the Claude key never reached the server. Now the app refuses to start without it and says why. - "UnknownHostException: smtp": MAIL_HOST was "smtp", not a real server name. Now a clear error; with no mail server, codes and alerts go to the log instead of crashing. 7) Bad input is stopped before it costs money Missing channel, unknown customer, empty message, bad date filter: readable 400/404 before any AI call. 8) Filters return exactly the right slice Phone + last 7 days showed only Phone rows from that week, in the table and the charts. Filters persist across pages. 9) Responsive design, pixel by pixel Broke and fixed: a score of 100 misaligned the score bar; mobile pagination wrapped into 4 lines; chart labels vanished on phones; the customer popup was clipped inside the table; the logo disappeared on deep links like /requests/12. Result: no horizontal scroll at 360 px. 10) A low score reaches a human "I'm cancelling my account" scored 8/100 and triggered an alert email to the admin's verified address, with score, reason, customer, assistant and a direct link. Verification codes expire in 10 min, allow 5 attempts, and have a 60 s resend cooldown. Today: an angry customer is found when a complaint escalates. We flag them within seconds. TODAY vs CLIENT PULSE AI - Coverage: small manual sample | 100% of conversations - Speed: days or weeks | seconds - Consistency: reviewer opinion | same rubric plus a written reason - Reaction: after the customer leaves | email alert below 20

Worth 20 of the 100 points.

Running it

SETUP — run Client Pulse AI locally Requirements: JDK 26, Node.js 22+, PostgreSQL 14+ (or Docker), and an Anthropic API key. 1) Get the code git clone https://github.com/jahangir-alisoy/client-pulse-ai.git cd client-pulse-ai 2) Start PostgreSQL (the app's default credentials) docker run -d --name clientpulse-db -p 5432:5432 -e POSTGRES_DB=clientpulseai -e POSTGRES_USER=clientpulseai -e POSTGRES_PASSWORD=clientpulseai123! postgres:16 (Or create the database "clientpulseai" with user "clientpulseai" / password "clientpulseai123!" in your own PostgreSQL.) 3) Start the backend (port 8080) cd client-pulse-ai-backend export ANTHROPIC_API_KEY=sk-ant-... (Windows PowerShell: $env:ANTHROPIC_API_KEY="sk-ant-...") ./gradlew bootRun (Windows: gradlew.bat bootRun; if "permission denied": sh ./gradlew bootRun) Tables and demo customers/support assistants are created automatically on first start. 4) Create a login user (once, after the backend has started) docker exec -it clientpulse-db psql -U clientpulseai -d clientpulseai -c "INSERT INTO users (username, password, api_key_provisioned) VALUES ('admin', '{bcrypt}\$2a\$10\$3YQbsGiZY8GbfxbYZAcbj.9rJNd1GVGW7styTJuqGiq6cQZIjn/gW', false);" Login: admin / secret123 5) Start the frontend (port 5173) cd ../client-pulse-ai-frontend npm install npm run dev Open http://localhost:5173 and sign in. 6) Try it - A default API key is created on first login (API keys page). - Simulation: pick a channel, customer and support assistant, then chat. Each turn is scored in the background. - Requests / Overview: scores, AI reasons, filters and charts. - Settings: change username or password, pick a theme, set an alert email. Optional email alerts (score below 20): set MAIL_HOST, MAIL_PORT, MAIL_USERNAME, MAIL_PASSWORD and MAIL_FROM before starting the backend. Without them, verification codes and alerts are printed to the backend console. Known limitation: AI scores need a valid Anthropic API key; without one, chat requests fail with an authentication error.

Project ownership

The team keeps the rights to this project. Judges were given access for verification only.