onboardiq

• 

  dashboard

💡 Inspiration

Business teams spend 2–4 hours manually turning raw CSV exports into something an executive can read. We asked: what if that entire process ran itself, autonomously, in under 60 seconds?

🏗 How We Built It

OnboardIQ is a 7-stage autonomous pipeline built on Next.js + NestJS, deployed on Alibaba Cloud ECS with Aliyun MaaS powering the AI layer.

The key architectural decision: stats-first, AI-second. We compute all statistics (Pearson/Spearman correlations, IQR outliers, missingness, categorical inconsistencies, date ambiguity) deterministically before any LLM is called. Then Qwen-Max interprets those hard numbers via a 6-function structured tool-calling schema — it cannot invent a finding that doesn't exist in the data. Qwen-Plus then synthesizes the final executive report. A human-in-the-loop gate holds the pipeline until an executive approves, rejects, or requests changes — which re-queues only the report stage using cached outputs.

🧗 Challenges

Getting the AI to be trustworthy without being rigid — designing the tool schema so Qwen could only reference real computed values — was the hardest problem. Engineering the pipeline to be resumable from any checkpoint without redundant reprocessing was the most complex infrastructure challenge.

📚 What We Learned

Structured tool calling with pre-computed context is the right pattern for reliable AI data analysis. Two-model chaining (one to interpret, one to write) produces significantly better output than a single model doing both.

Built With

• 10-redis
• 11-prisma
• 12-mysql
• 13-socket.io
• 14-node.js
• 15-nginx
• 16-pm2
• 19-svg
• 2-aliyun
• 24-rest
• 3-alibaba
• 5-nestjs
• 7-react
• 8-typescript
• 9-bullmq
• api
• authentication
• cloud
• ecs
• mysql
• qwen