M.Tech is not marked on working software. It is marked on a defensible contribution — a method, a comparison, a dataset, or a measurable improvement that did not exist before you started. A system that runs is the floor, not the deliverable. The question every panel is really asking is: what is new here, and how do you know it works?
So the ideas below are chosen for one property: each has a clear novelty angle you can turn into a paper. Every entry lists the question the panel opens with, because at this level that question is about your contribution, not your code. Before you pick, decide what your baseline is and what number you intend to move — that decision is your whole thesis.
Machine learning and deep learning
The most common M.Tech route, and the most crowded. The way out of the crowd is never the model — it is what you measure. Pick a baseline from the literature, beat it or explain honestly why you cannot, and report the failure cases. A results section that only contains wins reads as fabricated.
| Project | Stack | Difficulty |
|---|---|---|
| Recommendation system with a cold-start improvementThe novelty is your handling of new users or items — a measurable lift over a plain collaborative-filtering baseline is your contribution.Viva question it attracts: “What is your baseline, and by how much do you beat it?” | Python, scikit-learn or PyTorch, a public dataset | Hard |
| Fake news / misinformation detection with explainabilityThe contribution is not classifying — it is explaining why, and reporting where the model fails. That is what turns a demo into a paper.Viva question it attracts: “Show me an example your model gets confidently wrong.” | Python, transformers, LIME or SHAP | Hard |
| Resume screening / candidate ranking with bias analysisRanking is easy; auditing the ranking for bias is the research angle and the part a panel will push hardest on.Viva question it attracts: “How do you know your ranking is not biased by the wording?” | Python, NLP embeddings, a fairness metric | Hard |
| Plant disease detection with a lightweight / edge modelThe novelty is the trade-off — accuracy against model size or inference time — presented as a curve, not a single number.Viva question it attracts: “What did you give up in accuracy to run this on a phone?” | Python, CNNs, transfer learning, TensorFlow Lite | Hard |
| Face recognition attendance robust to spoofing or occlusionThe contribution is rejection — proving the system refuses a photo or an unenrolled face is harder and more publishable than matching.Viva question it attracts: “What happens when someone holds up a printed photo?” | Python, OpenCV, a face-embedding model | Very hard |
Applied systems and evaluation
Underused, and often the stronger thesis. Building a dataset, or measuring something everyone else only demos, is a genuine contribution — and it is far easier to defend than yet another classifier, because the novelty is the evaluation itself.
| Project | Stack | Difficulty |
|---|---|---|
| RAG-based domain assistant with a retrieval evaluationAnyone can wire up retrieval; the contribution is measuring retrieval quality and grounding, and showing where it hallucinates.Viva question it attracts: “How do you measure whether an answer is actually grounded?” | Python, a vector database, an LLM API, RAGAS or a custom eval | Hard |
| Comparative study of models on a domain dataset you builtA carefully built dataset plus an honest comparison is a legitimate M.Tech contribution on its own — the data is the novelty.Viva question it attracts: “Why is your dataset a fair test, and how did you split it?” | Python, several model families, a held-out test set | Hard |
| Anomaly or intrusion detection with a new feature setThe angle is the features you engineered and the precision-recall trade-off at a chosen operating point, not raw accuracy.Viva question it attracts: “At your threshold, what is the false-positive rate?” | Python, scikit-learn, a public security dataset | Hard |
What turns an M.Tech project into a thesis
- A named baseline. “My model gets 94%” means nothing without the number you are beating. Cite the prior work you compare against, and reproduce its result on your split.
- An honest limitations section. Every method fails somewhere. Finding your own failure cases before the panel does is the difference between a viva you control and one you do not.
- Reproducibility. Fixed seeds, a documented split, and a script that regenerates every number in your thesis. Examiners at this level ask to see it run.
- A paper-shaped write-up. Problem, related work, method, results, limitations. If your report already follows a paper’s structure, a submission is a formatting job, not a rewrite.
Every ML build on this site ships with the full source code, the trained model, the dataset pipeline and the evaluation scripts that produce your metrics — plus daily sessions that walk through the method and the results line by line, so you defend the contribution and not just the syntax. Start with the M.Tech projects. The research-angle builds above map to the recommendation system, fake news detection, resume screening, plant disease detection, the RAG college assistant and face recognition attendance. Each is capped per college, and the Mentored tier varies a module — the dataset or the domain — so your contribution is genuinely yours.
If your guide has already approved a specific novelty, have it built and evaluated for you. And frame the work early — the synopsis format is where you state the contribution your thesis will be graded on.