Hacktoberfest x MSC KBTCOE Nashik 2026

Developers often receive UI bug reports as screenshots accompanied by vague descriptions such as "the button looks wrong on my phone."

Build a multimodal tool that takes a screenshot together with a short description and identifies what may be wrong. It could turn the information into a structured bug report, suggest possible causes, or recommend where a developer should investigate.

You may also explore comparing screenshots from two versions of an interface and identifying meaningful visual changes.

Students working through proofs, circuits, diagrams, equations, designs, or handwritten solutions often receive feedback only after someone reviews their work.

Build a Socratic tutor that looks at a learner's work and helps them discover mistakes without immediately revealing the solution. Instead of solving the problem, the system should guide the learner through carefully chosen questions, observations, or small hints.

The challenge is not only understanding the work, but controlling how much assistance the model provides.

A single model response is not always reliable. A lightweight model makes it possible to approach the same problem from several perspectives.

Build a system where multiple Gemma 4 instances or roles independently examine the same image, text, or problem. For example, one could act as a skeptic, another as a domain expert, another as a beginner, and another as a verifier. The system should reconcile disagreements and produce a final response.

Prove whether it helps: Compare the multi-role approach against a single model call on a small test set. Report where multiple perspectives improve the result, and where they do not.

Long-running AI conversations eventually reach a point where everything can no longer remain inside the model's active context. Important goals, decisions, preferences, unresolved questions, and unfinished work may then disappear.

Build a system that preserves continuity across context boundaries. As the active context approaches a chosen limit, use Gemma 4 to create a compact handover containing the information a new session needs to continue effectively.

Explore challenges such as deciding when a handover should happen, deciding what information deserves to survive, keeping handovers compact without losing important detail, and maintaining accuracy across several successive handovers.

Testing idea: Use an intentionally small context budget during development so repeated handovers are easy to test.

Prove it: Place important facts or decisions early in a conversation, run through several handovers, and measure how well the system preserves them.

Fine-tuning is not always practical, but prompting, examples, retrieval, structured context, and self-checking can significantly affect model behaviour.

Choose a narrow and messy domain, such as handwritten records, receipts, laboratory conventions, club documentation, or a particular writing or review style. Build a system that improves Gemma 4's reliability on that task without fine-tuning the model, using techniques such as few-shot examples, retrieval, structured prompts, validation, and self-checking.

Measure the difference: Report performance before and after each technique so you can show which approaches actually helped.

Models are usually asked to answer. Make one act.

Build a game, simulation, AR experience, interactive installation, or simple camera-based system where Gemma 4 observes what is happening and decides what should happen next. Examples could include a game character that reacts to what a player draws, a camera system that decides which events are interesting, or an interactive environment that responds to visual input.

The challenge is not only making the idea work, but handling latency, unreliable outputs, unexpected situations, and what happens when the model makes a bad decision.