Glimpses of Agentic AI practises

Anushka Singh — Wed, 14 Jan 2026 18:28:51 +0000

In continuation of my Agentic AI learning I got to make a project complaint-triage-system by incorporating process automation with the help of LLM API. For database I used Sqlalchemy and when admin intends to check the lodged complaints on separate dashboard, they have to authenticate with their email and password secured by JWT token. It took me so long because first I used Gemini API key and at last I had to revoke because some glitch happened to be there (as usual). Oh I completely forgot to write that why did I use API key because it was helping me to triage the complaint status to high/medium/low and it was responsible for analysis of complaint submitted but the twist came when I wished to create the email send reply button by admin side along with regenerate and edit button.

I don't know how but my gemini key crashed, I rushed into changing the models but it didn't work. I switched to grok API because it was free, triage was working correctly and the analysis was a bit short (very specific in keywords).In between all of this It was all pain to use JWT based tokenization , I had to cancel otp based authentication when entering email, the otp was appearing to me in my backend server. I got so confused, my biggest red flag was lack of system design or clear workflow, and I mean it you not only need mere inspiration,but also clarity in your aim...

LET'S Get back to the topic so i just made a create_admin file and added my email, password in the env. After this, everything was working fine and I finally wanted AI-generated customized editable reply to be sent to the user.. I used APP password of one of my email ids but again it didn't work out. What came as my saviour Twilio Sendgrid. Sendgrid API key and my application was working without bugs, I edited my reply to add some human touch and specific details to the email however the edited reply was not going to the payload.. I debugged one last time and magic it is running smoothly.

Deploying is another pain and I tried on AWS Elastic Beanstalk but that http and https mismatch because I was using AWS amplify(using https), I turned to EC2 and tried installing nginx but the timeout in free tier and repeated commands exhausted me...This is how my full-stack application came into the life and the github commits> 11 has other memebase.
Today My aws free tier expired and I have multiple quests to create many projects, to contribute.
P.S. I intend AI to routing the complaint to specific department as per the user needs for faster issue resolving as my project grows

Whitepapers, Labs and loads of learning

Anushka Singh — Mon, 08 Dec 2025 18:47:25 +0000

This is a submission for the Google AI Agents Writing Challenge: Learning Reflections

Agentic AI is the evolution of automation which is agile in adaptive
decision making.

Fellas ! It was truly an amazing event packed with practical implementation

I started with Building your first Agent, it was quite fun and accomplishing to record the working of an agent. The learning is based on utilising ADK and Gemini for API keys and compatible model. As I geared up on day 1, I got to know the very basics of what is Agentic AI and the method it invokes to work with, then the types of agents and clear workflows of architecture in multi-agent systems.

I caught up with Interoperability with MCP and tools to get to know the deeper side of orchestration behind an agent's success. The best part was that you can connect the documents to the NotebookLM and learn the core concepts a way much better. They even provided the summary podcast created by NotebookLM.

I was able to call get tiny image tool from MCP server to test on my local host and it worked. Furthermore I wanted the output of image of an anime girl when asked from the agent just to work with different MCP server, let alone Replicate MCP server. There was a glitch and I moved on to day 3, however on day 2, I worked with agent with approval and definitely wouldn't have missed.
Day 3 and the dawn of context engineering Sessions and Memory one of my favourite topics, I put pen to paper and dived into how to make the agent stateful and the labs were my only resources cut to the chase for any beginner and I am glad everything was so smooth while I learned.

Day 4 was the addition to Responsible AI and if the agent is capable to solve problem Should it actually do or not, they must be evaluated too.
From Glassbox and Blackbox evaluations to pillars of observability. There was proper guidance depending on the roles of any professional.

While Day 5 was all about Prototype to Production , Deploying the agent.
On a good note, I would say the event was more than worthwhile
The curated playbooks and the steps to develop the agents on kaggle is the foundation to my next move which is developing Agentic AI project end to end.

Thank you for the day

DEV Community: Anushka Singh

Glimpses of Agentic AI practises

Whitepapers, Labs and loads of learning

Fellas ! It was truly an amazing event packed with practical implementation