Selecting an AI Internship Project as a JNTUH Student: Scope, Data and Evaluation

Pick the AI project that has data you can legally download today, a scope you can finish in about ten working days, a metric you can state in one sentence, and a result someone can see in under two minutes. Score every candidate idea on those four criteria with the rubric below, and drop anything that fails on data or evaluation, however exciting it sounds.

This guide is an independent resource for students of Jawaharlal Nehru Technological University Hyderabad (JNTUH) and its affiliated colleges. It is not issued by, or endorsed by, the university.

Why project choice decides the outcome of a short AI internship

A short internship rewards a narrow, finished project far more than an ambitious, half-built one. In a 14-day window you lose the first two days to setup and reading, and the last two to cleanup, documentation and a demo. That leaves roughly ten days of real work.

Most failed student AI projects fail for the same few reasons:

• The dataset turns out to be locked, unlabelled, tiny or full of errors, and the time is spent hunting for data instead of modelling.
• The scope is a “system” (a full chatbot for a hospital, a complete self-driving stack) rather than one question answered well.
• Nobody can say what “good” means, so the final report says “the model works well” with nothing to back it.
• The result lives in a notebook that only the author can run, so the reviewer or interviewer sees nothing.

The four rubric criteria map directly to these failure modes: data access, scope, evaluation clarity and demo value.

The project-selection rubric

Rate each candidate idea from 1 to 5 on each criterion, using the descriptions in the table. Multiply by the weight, then add up. The maximum is 100.

The weights add up to 20, so the maximum is 5 × 20 = 100.

How to read your total

• 80-100: Start. The idea is well matched to a two-week window.
• 60-79: Start only after you fix the lowest-scoring criterion. Usually that means shrinking scope or finding a cleaner dataset.
• Below 60: Replace the idea or cut it down to a smaller question.
• Any 1 on data access or evaluation: Stop regardless of total. These two cannot be repaired halfway through.

Criterion 1: Data access

Check data before anything else, because everything downstream depends on it. Spend one hour on this step and do not write model code during that hour.

Data checklist

1. Source: Where does the data come from? Public repositories such as Kaggle, the UCI Machine Learning Repository, Hugging Face Datasets, and Indian government open data portals (data.gov.in) are typical starting points.
2. Licence: Read the licence or terms of use. If you cannot find one, treat the dataset as unsuitable for a public demo.
3. Labels: Are the labels already present, and who produced them? Labels made by scraping or heuristics may be noisy.
4. Size: Can you load and train on it on a laptop or a free notebook environment? For a first project, tens of thousands of rows or a few thousand images is plenty.
5. Balance: Count the classes. If 97% of rows belong to one class, plan to use precision, recall or F1 rather than accuracy.
6. Personal data: Avoid datasets with names, phone numbers, faces of identifiable people or medical records unless they are published for research with clear permission.
7. Language: If you want Telugu or Hindi text, check that real labelled data exists. Many low-resource language tasks have far less data than English.

Be careful with data collected from your own college, a company or friends. It can look convenient, but it often needs approvals, anonymisation and consent that a 14-day project cannot absorb.

Criterion 2: Scope

Write your project as a single question: “Given X, can a model predict Y better than a simple baseline?” If you cannot phrase it that way, the scope is probably too wide.

Ways to shrink an idea

• Replace “a chatbot for students” with “a classifier that routes student questions into five categories”.
• Replace “crop disease detection system” with “image classification of leaves from three disease classes on one public dataset”.
• Replace “fake news detector” with “binary classification on one published dataset, with an error analysis on what it misses”.
• Use a pretrained model and fine-tune or prompt it rather than training from scratch.

A realistic ten-day plan

Criterion 3: Evaluation clarity

Evaluation is what separates a project from a tutorial copy. Decide the metric before you train anything, and write it down in your project note.

Match the metric to the task

Evaluation checklist

• Split the data into train and test before looking at test results; use a validation set or cross-validation for tuning.
• For time-based data, split by date, not randomly.
• Always include a baseline: majority class, mean prediction, or simple logistic regression.
• Check for leakage, such as duplicate rows or a feature that gives away the label.
• Report at least five real failure cases and a sentence on why each might have happened.
• Fix random seeds and note library versions so someone else can reproduce your numbers.

Criterion 4: Demo value

A demo is proof that the work exists. It also becomes something you can show in an interview or attach to a resume. It does not need to be fancy.

• Minimum: a notebook that runs top to bottom, with charts and sample predictions.
• Better: a small Streamlit or Gradio app where a person types a sentence or uploads an image and sees a prediction.
• Best: the app plus a README with a screenshot, setup steps, dataset link, metric table and a limitations section.

Keep a two-minute screen recording as a backup in case the app fails during a review.

Worked example: AI internship project (illustrative)

The following is an illustrative scenario, not a real student’s case, and no code or results are claimed. Suppose a third-year student wants to build “an AI to help farmers”. They list three candidate ideas and score them with the rubric.

Reading the scores: Idea A fails because labelled Telugu speech data is hard to obtain and the pipeline has many parts (speech recognition, understanding, response, speech output). Idea B is viable, but price data comes in irregular formats and forecasting needs a time-based split and a naive baseline. Idea C scores highest because the data has labels, the question is narrow and the metric is clear.

The student chooses C and writes the project question as: “Can a fine-tuned image classifier identify the disease class of a leaf image better than a simple baseline, measured by macro F1 on a held-out test set?” The plan is a baseline, a pretrained-model fine-tune, a confusion matrix, a review of misclassified images (for example, those taken in poor light), and an upload-and-predict demo. In the write-up the student also notes a limitation: photos from a controlled dataset may not match field photos taken on a phone.

Project ideas that tend to score well

• Text: sentiment or topic classification on a public review dataset; spam detection; classification of support tickets into categories.
• Vision: leaf disease classification; traffic sign recognition; mask or helmet detection on an open dataset.
• Tabular: loan default prediction on a public credit dataset; house price regression; student performance prediction from a published research dataset.
• Time series: electricity demand or air-quality forecasting with a naive baseline.
• Language models: a retrieval-based question answering demo over a public document set, evaluated on a small hand-made question list.

Ideas to avoid in a two-week window

• Anything that claims to diagnose illness or give financial advice to real people.
• Face recognition on classmates or staff.
• Training a large language model from scratch.
• Stock price “prediction” without a rigorous time-based evaluation. It is easy to produce misleading results.
• Projects that need robotics, drones or sensors you do not already have.

JNTUH regulations and choosing an AI internship project that fits them

JNTUH is a state university at Kukatpally, Hyderabad, and its R22 B.Tech regulations describe it as established under a Telangana State Act. It also works as an affiliating university: an undated excerpt from a JNTUH college site cites 423 affiliated colleges, a figure that may be outdated. Rules can therefore differ by regulation and by college. The JJR23 regulations of the autonomous constituent college at Jagtial show that autonomous units can publish their own rules.

For an AI project, this matters because of how graded internship work is structured. Under R22 B.Tech, students register for an Industry Training/Internship/Industry-Oriented Mini-Project/Skill Development Course after the II-Year II Semester exams. The work is done in the summer vacation or semester break and during III Year, and the report is presented to a committee in III-Year II Semester. The same regulations state that the course is evaluated for 100 external marks, and that a student fails with no report, no presentation or a score below 40. In practice, a two-week project with a clear scope, accessible data and a simple metric is easier to write up and present to an external examiner, so the rubric in this guide helps if you later want to discuss it with your department. Whether an online internship can count is not something those texts address, so do not assume it.

The final-year project follows a separate path. The R22 text describes a one-semester project beginning at the end of IV Year I Semester, with Project Stage I in that semester and an external examiner at the viva. Keep a short internship project distinct from this unless your department agrees otherwise.

Dates are another consideration. The academic calendars page lists B.Tech calendars, including 2026 entries, so check the current calendar for your year before planning a fixed 14-day window around exams or breaks.

The two-week internship at CareerSwami is free, runs online for 14 days, and includes an offer letter and completion certificate after successful completion. CareerSwami is independent and not affiliated with JNTUH. Treat it as extra practice, not a substitute for any college requirement.

Checks before you tie an AI project to JNTUH coursework

• Find your regulation (R22 or later) on your syllabus booklet and download the matching B.Tech regulations from jntuh.ac.in, since R22 text is dated 2022 and newer rules may apply.
• Read the internship/mini-project clause for your year: registration timing, committee composition and report format.
• Ask your HoD or project coordinator in writing whether an online internship can be shown for any requirement, and what documents they need.
• Confirm whether your college is affiliated or autonomous, because autonomous colleges may follow their own regulations.
• Check the current academic calendar for your year and note semester break dates before fixing your 14 days.
• If your project uses college or company data, ask who must approve its use.

Your one-page project note (template)

Before you start, fill in this template. If you cannot complete a line, your idea is not ready.

Project title:
Question (one sentence, "Given X, predict Y"):
Dataset name, link, licence, size:
Label source:
Baseline:
Primary metric and why:
Train/validation/test split method:
Planned model(s):
Planned improvement (one only):
Demo format:
Main risk and fallback dataset or model:
Rubric score (data / scope / evaluation / demo / total):

Common mistakes and quick fixes

Frequently Asked Questions

How many ideas should I score before choosing a project?

Score three to five. Fewer than three gives you no comparison, and more than five wastes time you need for building. Spend no more than one day on scoring.

Is a small dataset acceptable for an AI internship project?

Yes, if the question is narrow and you evaluate honestly. A few thousand labelled examples can support a good classification project, particularly with a pretrained model. State the dataset size and its limits in your write-up.

Should I use a pretrained model or train from scratch?

For two weeks, use a pretrained model and fine-tune it or use it as a feature extractor. Training from scratch needs more data and compute than most students have, and it leaves little time for evaluation and a demo.

Can I use ChatGPT-style tools or code assistants while building?

Check any rules set by the internship and by your college. If tools are permitted, use them to speed up routine code, but make sure you can explain every step of your pipeline, since you may be asked about it in a review or interview.

What if my chosen dataset has no licence information?

Treat it with caution. Look for the original publication or source page, or choose a dataset with a clear licence. If you still use it, do not redistribute the data in your repository; link to the source instead.

Do I need a web app for the demo?

No. A clean notebook with sample predictions is acceptable. A simple Streamlit or Gradio app does raise demo value in the rubric, and it makes your work easier to show to people who do not read code.

Can I turn my internship project into my final-year project?

That depends on your department’s rules for your regulation and college, and on whether the work is substantial enough. Ask your project coordinator in writing, and bring your one-page project note so the discussion is concrete.

What should I put in my resume from this project?

Write one or two lines: the question, the data, the model, the metric with the baseline comparison, and a link to the repository or demo. Include only results you actually produced, and never present a project as a deployed product if it only runs on your laptop.

Related Artificial Intelligence guides

• How JNTUK Students Can Spot Genuine AI Internship Learning Opportunities Versus Vague Offers
• AI Internship Guide for Osmania University Students: Building a Project Demonstration
• AI Internship Prerequisites for RGPV Students: Python, Data Handling and Model Basics
• All Artificial Intelligence internship guides