Glimpses of Agentic AI practises

Source: Dev.to

Complaint Triage System

In continuation of my Agentic AI learning, I built a project called complaint‑triage‑system. Users can submit complaints via email. The backend uses SQLAlchemy for the database, and administrators access a separate dashboard where they authenticate with email and password secured by a JWT token.

Initially I used a Gemini API key for triaging complaints into high/medium/low categories and for analyzing the submitted text. After encountering glitches, I revoked the key. I then switched to the Grok API (free) which handled triage correctly, though its analysis was very keyword‑specific.

Development Challenges

• JWT‑based tokenization proved cumbersome.
• OTP‑based authentication was abandoned because the OTP appeared only in the backend server logs.
• Lack of clear system design and workflow became a major red flag.

I created a create_admin script and stored the admin email and password in environment variables. After that, everything worked until I tried to send AI‑generated, customizable replies to users. An App password for my email didn’t work, but Twilio SendGrid saved the day. With a SendGrid API key, the application functioned without bugs. However, edited replies weren’t reaching the payload at first; after debugging, the issue was resolved and the system ran smoothly.

Deployment Hurdles

• AWS Elastic Beanstalk caused HTTP/HTTPS mismatches because I was using AWS Amplify (HTTPS).
• Switching to an EC2 instance with Nginx led to timeouts on the free tier and repeated command failures.

Despite these obstacles, the full‑stack application was completed and the GitHub repository now contains multiple commits.

My AWS free tier has expired, and I have many more projects to start and contribute to.