Final year project ideas for AIML: 22 with Indian data
APJ-RAISE team · · 11 min read
A good final year project for AIML or AI&DS students solves a real problem with real data, beats a simple baseline and has one part that is your own. A movie recommender or a Titanic survival model has none of that, and examiners have seen both hundreds of times.

This post is built around one question: what makes a final-year project count as research, and not only as a build? We answer it first, then give 22 project ideas on public data, most of it Indian. Each idea lists what you build, the data, a difficulty level, the part that can make it new and a repo to start from. We opened every link on 18 November 2026.
Final year project vs research project: what is the difference?
A build project shows that something works. A research project answers a question whose answer you did not know at the start. The same app can be either one, depending on what you measure.
- Build: "I made a crop disease app with high accuracy." The score comes from clean lab photos, so it says little.
- Research: "Accuracy falls from lab photos to field photos by this much, and this fix recovers part of the gap." Now there is a finding.
Most project lists stop at the build. They give a title and a dataset, and the student ends up reporting one accuracy figure on a test split that dozens of batches have used. To count as research, a project needs four things on top of the build:
- A baseline: a simple method you must beat, such as "same as last week" for prices.
- A fair test: one clear metric, measured on data the model never saw.
- Your own part: a new language, a new test set, a field test or an honest error analysis.
- Failures shown: examples where the model is wrong, with a reason.
A working demo still matters for the viva, since an examiner should be able to run it in two minutes. It just is not what makes the project research.
Worked example: from pothole app to research question
Start with a common idea: detect potholes in phone video. As a build, you train YOLO on RDD2022 and report mean average precision. Any batch can do that.
Now notice that RDD2022 has road images from six countries, India among them. That gives you a question: how much worse does the detector do on Indian roads if no Indian images are in training? Train twice, once with and once without the Indian images, keep every other setting fixed, and test both on video from your own city. Repeat with three random seeds. If the gap is large, try one fix, such as adding a few hundred labelled frames from your city, and measure how much of the gap closes.
The app is the same. The project is now a comparison with a clear answer, and you can defend every number in it.
How hard are these AIML project ideas?
We counted the difficulty labels across the 22 ideas below. Five are open to beginners (two beginner and three beginner to intermediate), eight are intermediate, and nine are intermediate to advanced or advanced. More than half also ask you to collect or label some data yourself, which is the easiest way to own a part of the work. If your team has one semester and no GPU, start with the beginner and intermediate ideas.
Agriculture and climate project ideas
India has rich open data for farming and weather. Most students never use it, which makes these projects stand out.
- Mandi prices: Build a model that forecasts the price of one crop in your district's markets two to four weeks ahead. Data: the daily mandi price table on data.gov.in, which comes from Agmarknet. It shows the current day, so save it daily and use Agmarknet reports for older prices. Level: intermediate. Novel angle: show honestly whether deep models beat a "same as last year" baseline, then add rainfall. Start with: statsforecast and neuralforecast.
- Crop disease: Build a phone-friendly leaf disease classifier that works on field photos, not only clean lab photos. Data: train on PlantVillage and test on PlantDoc, which has 2,598 real-world images of 13 plant species. Level: beginner to intermediate. Novel angle: measure the drop from lab to field photos, then shrink it with augmentation or fine-tuning. Start with: timm.
- Flood maps: Build a model that marks flood water in Sentinel-1 radar images, which see through clouds during the monsoon. Data: Sen1Floods11, about 14 GB of hand-labelled image tiles in a Google Cloud bucket. Level: advanced. Novel angle: run the trained model on a recent flood in an Indian state and check it against later cloud-free images. Start with: TorchGeo.
- Air quality: Build an hourly AQI forecaster for one monitoring station in your city. Data: the real-time AQI table on data.gov.in, from CPCB, with a free API key after login. It gives current values, so start logging early. Level: intermediate. Novel angle: pick a smaller city that few studies cover, and show which pollutant drives the forecast error. Start with: statsforecast.
- Crop mapping: Build a map of crop types in your district from Sentinel-2 satellite images. Data: Sentinel-2 through Google Earth Engine, plus 200 or more field points you label with a phone GPS. ISRO's Bhuvan portal has Indian map layers to compare with. Level: advanced. Novel angle: your own ground-truth labels. Start with: geemap.
Indian language and speech project ideas
AI4Bharat at IIT Madras and other Indian groups have released open models and data for Indian languages. These projects use them to serve people that English-only tools miss.
- Dialect ASR: Build speech-to-text for your district's dialect and measure where it fails. Data: IndicVoices (free, needs a Hugging Face login). Level: advanced. Novel angle: report word error rate by district and speaker group, not one average. Start with: IndicConformer, which covers all 22 scheduled languages.
- Lecture subtitles: Build a tool that adds English subtitles to recorded lectures by Indian teachers. Data: test on Svarah, with 9.6 hours of Indian-accented English from 117 speakers. Level: intermediate. Novel angle: add a glossary of technical terms and measure errors on those terms alone. Start with: Whisper.
- Notice translator: Build a translator for college or government notices into a low-resource language such as Santali, Bodo or Manipuri. Data: IN22-Gen for testing, plus 200 real notices you translate with a native speaker. Level: advanced. Novel angle: a small domain test set that did not exist before. Start with: IndicTrans2.
- Hinglish hate: Build a classifier for hateful posts in Hindi and Hindi-English code-mixed text. Data: HASOC 2021, with English, Hindi, Marathi and code-mixed sets (password through a Google Form). Level: intermediate. Novel angle: label 300 recent posts and test whether a model trained on 2021 data still works. Start with: IndicBERT.
- Sign language: Build a webcam app that recognises Indian Sign Language word signs. Data: INCLUDE from IIT Madras and AI4Bharat, with 4,287 videos of 263 word signs. Start with INCLUDE-50, a 50-sign subset. Level: intermediate to advanced. Novel angle: test on signers who are not in the dataset. Start with: MediaPipe hand keypoints.
- Sign reader: Build an app that reads shop signs and bus boards aloud in the local language, for people with low vision. Data: 300 photos you take in your town. Level: intermediate. Novel angle: a new local test set and an error analysis by script. Start with: IndicPhotoOCR (11 Indian languages and English) and Indic Parler-TTS for speech.
- Handwriting: Build a web page that recognises handwritten Devanagari characters in the browser, with no server. Data: the Devanagari Handwritten Character Dataset. Its page reports 98.47% test accuracy, so accuracy alone is not new. Level: beginner. Novel angle: collect characters from 30 classmates and measure the drop on real handwriting. Start with: ONNX Runtime Web.
- Textbook audio: Build a tool that turns a state-board textbook chapter into audio in its own language. Data: one public textbook chapter. Level: beginner to intermediate. Novel angle: a listening test where 20 people rate clarity, compared across two voices or two models. Start with: Indic Parler-TTS.
LLM and AI agent project ideas
LLM projects are popular, so most look the same. What sets yours apart is a test set you built and an honest count of failures. Read what AI agents are and our OWASP Top 10 for LLMs guide first.
- College RAG bot: Build a chatbot that answers questions from your university's rule book in English and Hindi, and says "not found" when the answer is missing. Data: your college's public handbook, plus 100 test questions you write. Level: intermediate. Novel angle: your test set and a measured rate of made-up answers. Start with: LlamaIndex and Ragas for scoring.
- Small model: Fine-tune a 1B to 3B model to sort student emails into the right department, and compare it with a large model. Data: 1,000 messages you collect and label with permission. Level: beginner to intermediate. Novel angle: cost and speed per message next to accuracy. Start with: Unsloth on a free Colab GPU. Our post on small language models explains the trade-offs.
- Legal summaries: Build a summariser for Indian Supreme Court judgments. Data: IN-Abs, judgments with summaries, in the Law-AI summarization repo. Level: advanced. Novel angle: ask law students to count claims in each summary that the judgment does not support. Start with: the repo's own baselines.
- Injection test: Attack your own RAG bot with prompt-injection and jailbreak probes, then add defences and measure again. Level: intermediate. Novel angle: a before-and-after table for one real app. Test only systems you own. Start with: garak, an open LLM vulnerability scanner.
- Voice price bot: Build a bot where a vendor asks "what is the onion price in Nashik?" by voice and hears the answer. Data: mandi prices as above. Level: intermediate to advanced. Novel angle: a small field test with real users, with their consent. Start with: IndicConformer for speech-to-text and Indic Parler-TTS for the reply.
Deep learning project ideas: vision, audio and science
- Potholes: Build a detector that finds potholes and cracks in phone video and pins them on a map. Data: RDD2022, 47,420 road images from six countries, including India, with four damage types. Level: intermediate. Novel angle: train with and without the Indian images and measure the gap on your own city's roads. Start with: Ultralytics YOLO.
- Voice deepfakes: Build a detector for fake speech, then test it on fake Hindi speech. Data: ASVspoof 2019 for training, plus Hindi samples you create with an open TTS model. Level: advanced. Novel angle: does a detector trained on English fakes catch Hindi ones? Use only your own voice or voices you have permission for.
- IPL win odds: Build a ball-by-ball win-probability model for IPL matches. Data: Cricsheet free ball-by-ball files. Level: beginner. Novel angle: check calibration (when the model says 70%, does the team win 70% of the time?) and explain each prediction. Start with: SHAP.
- PINN lab: Measure how a heated metal rod cools in your college physics lab, then fit a physics-informed neural network to your readings and compare it with the textbook equation. Level: advanced. Novel angle: real measured data, which most PINN projects lack. Start with: DeepXDE and our PINNs guide.
Machine learning projects with source code: how to use them
Every idea above links to an open-source repo. Use it as a baseline you build on and credit. Using a library or a published baseline is normal research practice; downloading a finished project and presenting it as your own is not.
Run the repo first and get its numbers. Then add the part that is yours, such as new data, a new language, a new test or a deployed app. In your report, write one line per repo saying what you used and what you changed.
Will examiners catch a copied GitHub project?
Often, yes. Examiners can search GitHub as easily as you can, and code-similarity tools exist. Stanford's Moss has compared student code for years. For written work, the UGC's 2018 plagiarism regulations for theses grade text similarity in levels: up to 10% carries no penalty, and penalties rise for 10 to 40%, 40 to 60% and above 60%. Your college may have its own rules for B.Tech projects.
- Commit often: a Git history built over months shows the work is yours.
- Credit every repo: name it in the report and in your README.
- Own one dataset: data you collected cannot be copied from anyone.
- Write it yourself: do not paste a repo's README or paper text into your report.
- Know every line: the viva will ask about code you did not write.
Can a final year project be published as a research paper?
Yes, when it answers a question rather than only building an app. The pothole example above is the pattern: one question, a fair comparison, failures shown. Two more steps get it ready for a venue.
- 1Share code and data on GitHub with a clear README, so others can rerun your comparison.
- 2Pick the venue carefully. A student workshop at a real conference beats a fee-charging "international conference" that accepts everything.
Our guide on how to start research as a student covers the full path. Before you pay any fee, read how to spot a predatory conference. If you want a research question rather than a build project, see our list of AI research topics for students. Our course on writing a research paper helps with the write-up.
At APJ-RAISE, students turn projects like these into papers in small teams, with feedback from peers. Joining is free for students aged 18 or older.
Questions people ask
- What are the best final year project ideas for AIML students?
- The best project uses real data, beats a clear baseline and adds one part of your own, such as a new Indian language or a test set you built. A crop disease detector tested on field photos is stronger than one tested only on lab photos.
- Can I use GitHub source code for my final year project?
- Yes, as a starting point that you credit. Name every repo you used and explain what you changed. Presenting someone else's finished project as your own is plagiarism, and examiners and code-similarity tools can find it.
- What are good deep learning projects for students?
- Good options include a pothole detector trained on RDD2022, an Indian Sign Language recogniser using the INCLUDE dataset, and a flood mapper using Sentinel-1 radar images. Each has public data and an open starter repo.
Sources
- 1.Singh et al. PlantDoc: A Dataset for Visual Plant Disease Detection. arXiv, November 2019.
- 2.Arya et al. RDD2022: A multi-national image dataset for automatic Road Damage Detection. arXiv, September 2022.
- 3.Javed et al. IndicVoices: Towards building an Inclusive Multilingual Speech Dataset for Indian Languages. arXiv, March 2024.
- 4.Javed et al. Svarah: Evaluating English ASR Systems on Indian Accents. arXiv, May 2023.
- 5.Gala et al. IndicTrans2: Towards High-Quality and Accessible Machine Translation Models for all 22 Scheduled Indian Languages. arXiv, May 2023.
- 6.Shukla et al. Legal Case Document Summarization: Extractive and Abstractive Methods and their Evaluation. arXiv, October 2022.
- 7.Sridhar et al. INCLUDE: A Large Scale Dataset for Indian Sign Language Recognition. ACM Multimedia 2020.
- 8.HASOC 2021 dataset page. FIRE 2021.
- 9.Open Government Data Platform India. Current Daily Price of Various Commodities from Various Markets (Mandi).
- 10.Open Government Data Platform India. Real time Air Quality Index from various locations.
- 11.Yamagishi et al. ASVspoof 2019: The 3rd Automatic Speaker Verification Spoofing and Countermeasures Challenge database. University of Edinburgh DataShare.
- 12.University Grants Commission. UGC (Promotion of Academic Integrity and Prevention of Plagiarism in Higher Educational Institutions) Regulations, 2018.
- 13.Aiken. Moss: A System for Detecting Software Similarity. Stanford University.

