- GenAI
- Projects
GenAI project ideas for students
The best GenAI project ideas for students are small apps you can measure: question answering over real documents with an evaluation set, an Indian-language voice bot, an offline translator and a structured data extractor. Most cost a few rupees to test, and the evaluation numbers are what make them worth showing in an interview.
By the DevIt teamPublished 5 min read
What makes a GenAI project worth showing
A chatbot that calls an API proves little, because anyone can build one in an afternoon. A project gets interesting when you can say how often it's right and what it costs per use.
So every idea below comes with two numbers to report: an accuracy figure from your own test questions, and a cost per request in rupees. If you can't measure a project, pick a smaller one you can.
Use public documents, not classmates' data. University rules, syllabus PDFs and government notices work well, and nobody's privacy is at risk if your app leaks them.
Five GenAI projects you can finish
Each one takes a known technique and adds the part tutorials skip. They're listed in the order that reuses the most code.
| Project | Technique | What to report |
|---|---|---|
| Ask your university's rulebook | Retrieval-augmented generation (RAG) with citations | Share of 50 test questions answered correctly |
| Test suite for the rulebook bot | Retrieval hit-rate, faithfulness checks | Hit-rate at top 3 chunks, wrong-answer examples |
| Hindi or Tamil voice FAQ bot for a college office | Speech-to-text, LLM, text-to-speech | Cost per call, word error on 20 recordings |
| Offline notice translator | Local translation model | Time per page on your laptop, errors a native speaker found |
| Receipt to JSON extractor | Structured output with schema validation | Share of receipts that pass validation first time |
Ask your university's rulebook, with citations
Load your university's regulations or a syllabus PDF, split it into chunks, embed them and answer questions by retrieving the closest chunks. Make the model quote the clause it used, with a page number.
Then write 50 questions a student would ask, such as how many attendance shortages are allowed or when re-evaluation fees are due, and record the correct answer for each. That list turns a demo into a project with a score.
The honest result is rarely perfect. Report the score anyway, and show two questions it got wrong and why.
What these projects cost in rupees
Price your project before you build it. Sarvam AI's API pricing is listed in rupees: its 105B chat model costs ₹15 per million input tokens and ₹60 per million output tokens on the pay-as-you-go tier, and new accounts get ₹100 of free credits.
A RAG question that sends about 3,000 tokens of retrieved text and gets a 300-token answer costs about six paise:
# Sarvam pay-as-you-go prices in rupees, checked 15 November 2026
INPUT_PER_MILLION = 15.0 # ₹ per 1M input tokens
OUTPUT_PER_MILLION = 60.0 # ₹ per 1M output tokens
def cost_inr(input_tokens: int, output_tokens: int) -> float:
return (input_tokens * INPUT_PER_MILLION + output_tokens * OUTPUT_PER_MILLION) / 1_000_000
per_question = cost_inr(3_000, 300)
print(f"₹{per_question:.3f} per question, ₹{per_question * 50:.2f} for 50 test questions")
# ₹0.063 per question, ₹3.15 for 50 test questionsAt that rate the ₹100 signup credit covers roughly 1,500 questions. Google's Gemini API also has a free tier "for developers and small projects getting started", according to its pricing page, so you can build most of these without paying anything.
Build for Indian languages
An Indian-language project is harder to copy from a tutorial, and it's useful to people who aren't programmers. A voice bot that answers a college office's common questions in Hindi or Tamil is a good start.
Sarvam lists speech-to-text at ₹0.49 a minute and its Bulbul V3 text-to-speech at ₹3 per 1,000 characters. A one-minute question with a 400-character spoken answer therefore costs about ₹1.69 for speech, plus a few paise for the model.
For offline work, AI4Bharat's IndicTrans2 is an open-source translation model for all 22 scheduled Indian languages, with code and checkpoints under the MIT licence. Run it on your laptop to translate notices, then ask a native speaker to mark the errors, and put their corrections in your README.
Turn receipts into clean JSON
Structured extraction is less flashy than a chatbot and closer to what businesses automate. Give the model a photo or PDF of a bill and ask for the shop name, date, GSTIN, line items and total as JSON that matches a schema you define.
The model's output is only half the project. Validate every response against the schema, reject anything that fails, and check that the line items add up to the total. A receipt where the maths doesn't match is a wrong answer, however tidy the JSON looks.
Collect 30 receipts of your own, from groceries, the canteen and online orders, and remove anything personal. Report three numbers in the README: the share that pass validation first time, the share that pass after one retry with the error message fed back, and the share where the total is right.
Measure it before you call it done
Evaluation is what separates a GenAI project from a demo. Two numbers cover most RAG projects.
Retrieval hit-rate is the share of test questions where the chunk containing the answer appears in your top 3 results. Faithfulness checks whether each answer is supported by the retrieved text, rather than invented by the model.
You can use a second model as a judge for faithfulness, but check 20 of its verdicts by hand and report how often you agreed with it. An evaluation you haven't checked is just another unmeasured model.
Building these with DevIt's GenAI internship
If you want graded projects and deadlines, DevIt's Generative AI Engineering Internship covers this ground. It expects working Python, a terminal, Git and a free-tier API key.
The 1-month plan includes a prompt engineering playground, a structured data extractor and a semantic search engine. The 3-month plan adds RAG over your own documents, a deployed RAG chatbot, a tool-using agent, an LLM evaluation project, a fine-tuning project with a before-and-after evaluation and a guardrails layer.
It's a paid training internship with no stipend: on 15 November 2026 the 1-month plan cost ₹199 at its Early Bird price. The internships page lists every track. For the backend side of deploying these apps, see backend projects for beginners.
Frequently asked questions
Do I need a GPU for GenAI projects?
Not for most of them. Calling a hosted model through an API runs on any laptop. You only need a GPU, or a free cloud notebook, for fine-tuning or for running larger open models locally.
Which GenAI project is best for a resume?
A RAG app with your own evaluation set. It shows retrieval, prompting and measurement in one project, and the accuracy score gives an interviewer something concrete to ask about.
How much does it cost to build a GenAI project?
Often nothing. Sarvam gives ₹100 of free credits on signup and Gemini has a free tier. At Sarvam's listed prices, a typical RAG question costs about six paise.
Should I fine-tune a model for my project?
Only after a RAG version with an evaluation set is working. Fine-tuning is worth showing when you can report the score before and after, on the same test questions.
Related reading
- Backend projects for beginnersBackend projects for beginners that show real skills: a REST API on PostgreSQL, safe logins, a Razorpay webhook, rate limits and a job queue.
- Frontend projects for your resumeFrontend projects for your resume should load fast on a phone and handle real data. Five builds worth your time, and how to host them for free.
API prices, free tiers and DevIt track details checked 15 November 2026. Model prices change often; recheck before you budget.