AI Enterprise SolutionsGPT & Claude

Mindrift CRM

THE USER AND THE PROBLEM Target Users: Customer support managers, sales representatives, account managers, and operations teams in small and medium-sized businesses that handle multiple customer requests across email, WhatsApp, contracts, and internal communication channels. The Problem: These teams often struggle to keep customer information, promises, agreed terms, and unresolved issues consistent across different communication channels. Important details get lost in long conversations, different employees may provide conflicting information about prices or deadlines, and customers are sometimes asked the same questions repeatedly when a case is transferred to another team member. As a result, businesses experience delayed resolutions, repeated work, missed follow-ups, customer frustration, and preventable customer churn. Employees also spend valuable time manually reviewing conversations and tracking unresolved issues instead of focusing on solving customer problems. The Core Pain Point: Businesses lack a reliable, centralized way to connect what customers requested, what the company promised, what actions were taken, and whether the customer's problem was actually resolved.

Open the demo Video GitHub

Preliminary score

67.0/ 100Promisingaverage of 2 judges

Value for the user18.5 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI22.0 / 30
A working core scenario, and what AI actually contributes to the result.
Quality testing10.0 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility9.5 / 15
Data requirements, running costs and a clear next step.
Originality7.0 / 10
A meaningful difference from common solutions.

What the judges wrote

Claude Opus 5.5Preliminary65/100
Strengths
Your evidence-first design is the standout: every AI finding must cite quotes that the backend actually finds in the source text, and unverified claims are downgraded rather than shown as fact. Your build is unusually complete for one day. It has Zod-validated structured outputs, hash-based re-analysis that avoids duplicate findings, deterministic guards against false contradictions (different packages, later-agreed discounts), role-based access, and WhatsApp and Telegram intake. ProofClose is a well-reasoned business rule that separates 'agent replied' from 'problem solved' and blocks case closure until the customer confirms.
To improve
Your deck admits that no test results or observed failures were recorded, and the manual-vs-Mindrift comparison is only a plan. Running even 10–20 labelled scenarios and reporting correct findings, false alerts and missed issues would greatly strengthen your case. The track asks for a metric that justifies adoption, so name one with a target, such as handover time or the share of cases reopened after closure. Make clear how much of your demo runs on live Claude versus pre-seeded 'ready results', and add concrete per-case model and hosting cost estimates. Several slides still contain unfinished template text ('must be confirmed before submission'); remove it before you pitch.
GPT-6 AstraPreliminary69/100
Strengths
Your case-centred workflow addresses a concrete B2B problem: commitments and unresolved issues getting lost between sales, delivery and support. Your source-linked findings, human approval steps and customer-confirmation gate provide thoughtful safeguards against unsupported conclusions and premature closure. Your README gives substantial implementation detail and transparently distinguishes prepared demo results, rule-based analysis and live AI.
To improve
Your next demonstration should process an unfamiliar case end to end in Real AI mode, showing ingestion, grounded findings, handover and confirmation-gated closure. Your testing report should include expected versus actual results for package differences, accepted discounts, duplicate messages and unauthorized access; the deck currently supplies scenarios but no observed outcomes or manual-review comparison. Your feasibility plan should quantify model cost per case and expected monthly operating costs, then define measurable success criteria for the proposed pilot.

The team

1)Ataxan Hacızadə -Frontend Developer
developer
Nihad ZəkiyevLead
product
2)Elcan Məmmədov -Backend Developer
developer

Disclosure

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

Models: Claude Opus 5.5 (claude-opus-5-5, Anthropic API). Analysis of customer cases: need-vs-proposal fit, cross-source contradictions, blockers, repeat problems, lost-deal analysis, handover packages, resolution assessment. Text extraction from camera photos of documents (vision OCR). - Tesseract OCR (tesseract.js, LSTM model, Azerbaijani and English language data). Offline in-browser OCR when no API key is configured. - Rule-based analysis (not ML). Limited fallback when AI is unavailable; results are clearly labelled "not AI". - Development tool: the AI coding assistant Claude Code (Anthropic) was used while building the project.
Data: - Synthetic data. All demo data is fictional: 6 Azerbaijani customer companies (Bakı Retail, Gəncə Mebel, Xəzər Logistika, Şəki Turizm, Lənkəran Aqro, Sumqayıt Aptek) with their cases, conversations and proposals. No real customer data is used. - User-provided data at runtime: texts, files (TXT/PDF), camera photos, WhatsApp and Telegram messages. - Open data: Tesseract language data (aze, eng, Apache 2.0 license).
Components: - Frontend: React 18, React Router, TanStack Query, Tailwind CSS v4, Vite, Lucide icons, tesseract.js. - Backend: Node.js, Express, Prisma ORM, SQLite, Zod, Anthropic SDK, unpdf (PDF text extraction), Multer, Helmet, bcryptjs, express-rate-limit. - Integrations: WhatsApp Business Cloud API (Meta), Telegram Bot API. - Testing & quality: Vitest, Supertest, TypeScript, ESLint. - Platforms: Render (hosting), GitHub (source code).
Built during the event: Confirmed

Quality testing

While building Mindrift CRM, we focused on a simple question: Can users trust the information the system gives them? We tested real-world scenarios where small mistakes could lead to bigger business problems. When we identified issues, we improved the relevant logic and re-tested the affected scenarios. AI accuracy: We checked whether AI-generated findings matched the original customer messages, so users could verify conclusions against their sources. False contradictions: We addressed cases where different prices for different service packages could be incorrectly flagged as contradictions. Duplicate messages: We tested duplicate WhatsApp messages and improved their handling to help keep customer case histories consistent. Case closure: We checked that cases could not be closed before the required customer confirmation. Access control: We reviewed permissions to prevent users from accessing cases they were not authorized to view. Messaging integrations: We tested the WhatsApp integration and our Telegram bot to support the flow of customer messages into the CRM. We also looked beyond individual features. We considered how these parts work together, because a message that is received but handled incorrectly can be just as problematic as a feature that does not work at all. Mindrift CRM is still a prototype, and we know that AI can make mistakes when information is incomplete or lacks context. That is why we focus on making findings verifiable, keeping business rules consistent, and handling customer information responsibly. For us, quality is not just about making the system work. It is about making sure it behaves as expected when real users depend on it.

Worth 20 of the 100 points.

Running it

SETUP INSTRUCTIONS 1. Install and Run: - Run npm run setup (first time only) - Run npm run dev and open http://localhost:5173 2. Login Credentials: - Use the demo login button or email with password: Demo123! 3. AI API Connection & Configuration: - Works without an API key in Demo mode, but a real AI API key can also be connected. - To enable Real AI, set ANTHROPIC_API_KEY in the server .env file and set DEMO_MODE="false".

Project ownership

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