From "What Is Agentic AI?" to Deploying AI Agents on IBM Cloud — My First Month as a Student Founder in a Tech Internship
I'm writing this at the end of Week 2 of a four-week IBM SkillsBuild internship, with my laptop open on three tabs — the watsonx Orchestrate agent builder, a half-finished architecture diagram for my final project, and this draft. I didn't plan to write this mid-internship. But something about the last two weeks felt worth documenting before the momentum of Week 3 erases the texture of it.
Who I Am, and Why This Internship Felt Different
I'm a first-year B.Tech CSE student at Government Engineering College, Buxar, Bihar. I'm also the founder of Mediokart, a rural healthcare startup — and before anyone asks, yes, I started building it before I joined college. The startup began as AuraBox, a smart IoT-based first-aid box, and has since evolved into Sahayak, a WhatsApp-based AI health assistant in Hindi and Bhojpuri. I run it with a mix of real mentors, online communities, and — not going to lie — a lot of conversations with AI tools to fill the gaps in my very small founding team.
This is my second internship with Edunet Foundation. The first was a Data Analytics with LLMs program in late 2025. So when I got the offer letter for this one — the IBM SkillsBuild four-week program on AI and Cloud Technologies, running from June 12 to July 10, 2026 — I wasn't starting from zero. But I was about to learn things that no previous experience had quite prepared me for.
Week 1: The Day "Agentic AI" Stopped Being a Buzzword
The first session was an orientation. Standard stuff. But Day 2 of Week 1 is where things shifted — we went deep into Agentic AI and got our first hands-on session with Langflow and IBM watsonx.
Here's the thing about agentic AI that I didn't fully get until I built something: it's not just a chatbot with a longer context window. An AI agent doesn't just answer questions — it decides what tool to use, when to use it, sequences multiple steps, and arrives at an answer through something that genuinely looks like reasoning. When I built my first RAG-based agent that week — uploading a document, connecting it to watsonx.ai via IBM Granite, and watching the agent retrieve grounded answers instead of hallucinating — something clicked.
RAG (Retrieval-Augmented Generation) is one of those concepts that sounds impressive until you understand it, at which point it sounds inevitable. You have an LLM that generates text, but it doesn't know your specific data. RAG gives it a memory — by retrieving relevant chunks from your uploaded documents and feeding those as context before generation. The LLM isn't guessing anymore; it's reasoning on verified information. This is going to be the backbone of my final project, and I'll come back to that.
That same week, I set up my Jupyter Notebook environment on Watson Studio, ran my first Python cells on IBM's cloud infrastructure, and spent an hour in the Prompt Lab — deliberately breaking prompts to understand what temperature, max tokens, and few-shot examples actually do to LLM outputs. This is the kind of hands-on intuition that you can't get from reading a paper.''
The IBM Granite Discovery
On my own time, separate from the internship sessions, I went into the IBM Granite model documentation. What I found was more than I expected — Granite models are IBM's own family of open-source LLMs (including granite-3-2-8b-instruct, which we used in the RAG demo), built with a strong focus on enterprise use and responsible AI. Understanding the model I was using, rather than treating it as a black box, changed how I thought about prompting and fine-tuning.
Week 2: The AutoAI Revelation (and a Cybersecurity Surprise)
Week 2 started with badge courses — Getting Started with Artificial Intelligence and Getting Started with Cybersecurity. I'll be honest: I went into the AI course expecting it to feel like a review, and into the Cybersecurity course expecting to check a box.
Neither happened the way I expected.
The AI course had a section on the evolution of AI that laid out the entire arc in a way I'd never seen condensed before — from ELIZA, the 1966 chatbot that mimicked a therapist using simple pattern matching, through the symbolic AI era, the expert systems phase, the machine learning wave, and finally to the generative AI moment we're living in now, with GPT-4, Gemini, and IBM's own Granite series. Seeing this as a timeline made today's "agentic AI" revolution feel like the logical next chapter rather than a sudden leap.
And then the Cybersecurity course hit differently than I expected. I went in thinking it was a checkbox. I came out with a new question mark over my own career path. The role of AI in cybersecurity — threat detection, anomaly spotting, phishing identification — is not a niche. It's becoming infrastructure. I'm now seriously thinking about this field in a way I wasn't before I opened that course.
AutoAI: What Happens When ML Becomes Accessible
The most technically dense and satisfying part of Week 2 was AutoAI on Watson Machine Learning.
Normally, building a machine learning model means choosing an algorithm, cleaning data, engineering features, hyperparameter tuning, cross-validation — hours of work for something that may or may not generalize. AutoAI does all of this for you. You provide a dataset and tell it what column to predict. It then runs experiments across multiple algorithms (in my case, it tried an Ensembler, Holt-Winters, and BATS for a time-series prediction task), runs each through a pipeline of hyperparameter optimization, feature engineering, and ensembling, and presents a ranked leaderboard.
I trained a Microsoft stock price prediction model using AutoAI, with the close price as the target variable. The winning pipeline — Pipeline 12, an Ensembler — achieved a SMAPE of 0.270 on validation and 0.250 on backtest. Then I deployed it as a live online deployment on Watson Machine Learning's scoring endpoint. The deployment space showed "1 Deployed, 0 Failed" — and when I called the endpoint via the browser's network inspector, I could see the JSON response coming back with actual predictions.
That moment — watching a model I'd trained answer a real API call — is different from seeing it work in a notebook. The model was no longer a file on my computer. It was a service.
I also ran a crop recommendation classifier as a side experiment that week. Multiclass classification, Snap Random Forest algorithm, 100% confidence on "mungbean" for the test input. Small experiment, but it completed the loop: understand AutoAI on time-series, verify it works on classification too.
Deploying Two AI Agents
Outside of the AutoAI focus, I also built and iterated on an AI agent for students and researchers on Watsonx Orchestrate — designed to help people discover hackathons, internships, and project ideas, compare frameworks, and generate project proposals. Skills/actions connected to web search, knowledge base, IBM watsonx.ai models. This was a natural extension of my own problem: I spend significant time hunting for these opportunities, and a conversational agent that centralizes discovery would have saved me real hours.
Going Off-Script: Quantum Computing and Speech-to-Text
Nobody assigned me these. I explored them anyway.
IBM Quantum Cloud — I spent an evening on quantum.cloud.ibm.com going through the introductory quantum computing course. I'm not going to pretend I understand quantum error correction yet. But I understand what a qubit is, why superposition and entanglement are not just metaphors, and where quantum computing is likely to intersect with AI in the next decade (specifically in optimization problems and cryptography). The IBM quantum computer visualization — the actual dilution refrigerator structure — is one of the most striking images I've encountered in a technical context.
Watson Speech-to-Text — I provisioned this as an independent IBM Cloud service (it's active in my account under Sydney region now). The reason is practical and tied to my final project: most ASHA workers in rural Bihar are far more comfortable speaking than typing. A voice-first interface could be the difference between a tool they use daily and a tool they open once and forget. I'm planning to integrate Speech-to-Text into the agentic workflow — not as a demo feature, but as a genuine accessibility unlock.
The Final Project: ASHA Worker Co-Pilot
This is where everything converges.
ASHA (Accredited Social Health Activists) workers are India's frontline community health workers — over 1 million of them, each responsible for roughly 1,000 people in their assigned habitation. They are the first and often only healthcare contact for rural families across India. Their work contributed significantly to India's maternal mortality ratio dropping from 130 per 100,000 live births in 2014–16 to 97 in 2018–20.
And yet their daily operational reality is brutal: they currently juggle 7+ separate apps that don't communicate with each other, maintain 10+ paper registers alongside digital entries, and operate in low-connectivity environments. In December 2024, Bihar discovered that health data for nearly 20 million people had not been uploaded — officials traced this to inadequate training and technical support, not negligence. A 2025 government survey found 64% of rural women in India cannot perform basic smartphone tasks — and yet ASHA workers are expected to operate multiple complex government platforms.
This is the gap I'm building into.
ASHA Worker Co-Pilot is a multi-agent AI system built on watsonx Orchestrate and powered by IBM Granite. It consolidates three of the highest-friction daily tasks for ASHA workers into a single conversational interface in Hindi/Bhojpuri:
Agent 1 — Triage Agent: Assesses symptom urgency (HIGH/MEDIUM/LOW) and recommends ANM/PHC referral for anything above LOW. Critically, it never diagnoses — it flags and refers. Responsible AI is not a compliance checkbox here; it's a design principle.
Agent 2 — Eligibility Agent: Matches a person's details (age, BPL status, widow/disability status, etc.) to the correct NSAP scheme — IGNOAPS, IGNWPS, IGNDPS, NFBS, or Annapurna — and tells the worker the benefit amount and where to apply.
Agent 3 — Risk-flag Agent: Checks 10 maternal health indicators (ANC visit count, age, anemia signs, previous pregnancy complications, etc.) against WHO/NHM guidance to flag high-risk pregnancies for immediate referral.
A Supervisor Agent routes every incoming query to the right specialist agent automatically — the ASHA worker just types (or eventually speaks) in natural language.
The knowledge base is grounded in real government documents: the NHM ASHA Facilitator Handbook, NSAP eligibility criteria, SDG 3.1 maternal health indicator data (the same dataset that appears in IBM's own AICTE problem statements). This grounding is what makes RAG essential here, not optional — I cannot have an agent recommending scheme benefits from outdated or hallucinated information.
The future version integrates Watson Speech-to-Text for voice input. The long-term roadmap includes WhatsApp Business API deployment (ASHA workers already use WhatsApp daily), integration with ABDM (Ayushman Bharat Digital Mission) for longitudinal records, and expansion to cover Bihar state-specific schemes beyond NSAP.
What Two Weeks Actually Taught Me
On learning: The fastest way to understand a technology is to build something with it where the output matters to a real person — not just to you. The moment I started designing the ASHA Co-Pilot with actual ASHA workers' pain points in mind, every technical decision (which agent for which task, how to write the instructions, what goes in the knowledge base) became purposeful. Compare that to building a practice project with dummy data — the learning is shallower because the stakes don't feel real.
On IBM Cloud tools specifically: Watson Studio, Prompt Lab, AutoAI, Watson Machine Learning, Watsonx Orchestrate, and Langflow are not just enterprise wrappers on existing open-source tools. They represent a specific philosophy — that AI development should be accessible without being superficial, that deployment should be as straightforward as training, and that enterprise-grade MLOps shouldn't require a DevOps team. As a solo founder, this matters.
On Cybersecurity: I came in not caring about it. I'm leaving Week 2 genuinely considering it as a parallel track to healthcare AI. The course showed me how AI is reshaping threat detection, social engineering defense, and anomaly identification — and honestly, securing Mediokart's patient-data infrastructure is something I should have been thinking about much earlier.
On Quantum Computing: Still early for me. But I understand now that this is not science fiction — it's engineering in progress, with IBM at the frontier. The timeline for quantum advantage in practical optimization problems is closer than most people in the software space assume.
What's Coming in Week 3 and 4
Week 3 brings the session on Generative AI for Cybersecurity, an introduction to Data Analytics, and the hands-on build of a phishing detection agent using watsonx Orchestrate. This is exactly where my new interest in cybersecurity and my existing agent-building skills intersect — I'm genuinely curious to see how far the phishing detection agent can go with real-world email data patterns.
Week 4 brings a college admissions chatbot build on watsonx Assistant, an introduction to Quantum Computing, a final AMA session, and — most importantly — project submission.
The ASHA Worker Co-Pilot will be ready by then. Not production-ready (four weeks is not enough time to safely deploy anything in a healthcare context), but proof-of-concept ready — a working demo that shows the full agent routing, RAG-grounded responses, and a live conversation in the kind of language a real ASHA worker in Bihar would actually use.
Final Thought
I started this internship already having done one before, already knowing some of the tools, already having a startup in healthcare AI. And I'm still surprised by how much I didn't know.
That's the thing about actually building — it reveals the gap between knowing something and understanding it. I knew what RAG was. I didn't understand it until I watched my own agent retrieve a NSAP eligibility criterion from a government document and answer a question that I'd verified manually was correct.
There's a version of this internship where you complete the assignments, collect the badges, and write a LinkedIn post. That's fine. But the version I'm in — where you use the internship as structured infrastructure to build something that's actually trying to solve a real problem — is different. It's harder and it's slower, and some days you feel like you're debugging the wrong thing entirely. But it's also the only version that teaches you something you can't unlearn.
ASHA Worker Co-Pilot ships in two weeks. Watch this space.
Aman Kumar Happy is a first-year B.Tech CSE student at Government Engineering College, Buxar, and the founder of Mediokart, a rural healthtech startup. He is currently completing the IBM SkillsBuild x Edunet Foundation AICTE 4-week internship on AI and Cloud Technologies (June–July 2026).
Connect on LinkedIn | GitHub: github.com/amankumarhappy
Comments
Post a Comment