- Event
- WolfHacks 2026 · 24 hours
- My role
- Interface vision, UX direction, modeling, full-stack development
- Team
- Srujan Ponnapalli, Rehan Mohammed, Sampath Kasina, Aadee Chheda
- Built with
- Next.js, TypeScript, Python, scikit-learn, Tiger Cloud, Gemini
1st PlaceWolfhacks 2026
pathway
From a warning to a workable week.
I helped shape Pathway around a simple question: once a student learns a course needs attention, how do they see why and decide what to do about it?
An explainable student-support prototype, built in 24 hours.
What should I focus on before the BIO 181 midterm?
A risk label can create urgency. It can't make a plan.
An early alert says a class may be slipping. It does not show which observations produced that signal, which habits a student can change, or where a response fits before the next deadline. A prediction on its own leaves the student with the hardest work.
I defined the interface around a continuous route: notice a course worth checking, understand the model's reason, then act on a feasible change. That idea connected Home, Insights, Week, and Viva instead of presenting them as isolated features.
Works with everything.
Course records, files, the calendar, mail, and sleep: Pathway starts from the tools a student already uses. In the prototype that means uploads and one-click handoffs; deeper connections are planned.
Follow BIO 181 from check-in to a place in the week.
The sample semester gives the story a concrete subject. A BIO 181 midterm is approaching; the student can move from a course-level alert to an inspectable reason, a hypothetical change, and a proposed study block. The same course context persists at each step.
The signal sits beside the week.
In the synthetic BIO 181 semester, Home marks the course worth a check-in before a midterm. The next task sits near the student's schedule, so the alert starts a route instead of ending at a risk label.
We made the course, rather than an abstract score, the starting point for the experience.
Midterm in four days
Course facts from sample syllabi and grade pages retain source lines. Viva can use that context to propose a calendar edit, but the edit waits for the student's approval. A what-if changes a hypothetical model result; it does not change the student's actual standing.
Choose a model the student can inspect.
We built two small decision trees: an early tree before a midterm score exists and another for mid-semester. Python trains and exports them; TypeScript walks the exported tree in the browser. Insights can show the actual decision path and search for a supported change that crosses a different model leaf.
What the tree caught. What it missed.
160 held-out synthetic records · 44 finished below C
- Recall
- 73% of below-C records flagged
- Precision
- 63% of flags matched a below-C record
The early tree missed its 80% recall target by 7 percentage points on this held-out set.
Readability had a measurable cost.
The early tree missed its 80% recall target on held-out data. An early logistic-regression reference caught 36 of 44 below-C outcomes with 18 false alarms, outperforming the tree on some measures. We retained the tree because a student could follow its rules and try a constrained what-if. The test set was viewed twice during revision; both looks are reported.
Keep the student in the decision.
The model can indicate a boundary. The interface has to explain what that boundary means and leave control with the person living the week.
Call it a check-in, not a verdict.
“Worth a check-in” creates a reason to look closer without claiming the model knows the student's future. The product shows the path and the similar synthetic records behind the result.
Separate a what-if from reality.
Students can test supported habits such as earlier exam preparation. The result is a model scenario, not a promise of a higher grade or a change to their real record. Context such as work hours and a midterm score is never prescribed as a behavior to change.
Ask before changing the week.
Viva carries course context into a proposal, with a dated action and its source visible. A calendar change waits for approval, and the student can review the result in Week.
One course context across the product.
Beyond the model, I defined the interface vision and built major parts of the app shell, Home and Insights flows, model evidence, landing page, and Tiger Cloud persistence. The team connected course materials, study tools, Viva's chat and voice modes, calendar actions, an inbox, and study buddies.
The tree runs locally in the browser. Course records, activity, plans, and chats persist when the database is configured. Gemini powers optional document analysis and assistant behavior; ElevenLabs handles speech when configured. The core sample demo works without those services.
A source-based quiz can tie feedback to a lecture file. Completing a block alone does not guarantee a new course status: the model re-scores only when an actual input changes. That distinction was central to keeping the product honest.
The wider prototype
- BeaconA Raspberry Pi desk view keeps the course signal, next step, and Viva nearby.
- Study PotsA Solana devnet experiment uses valueless test tokens for a shared study plan. Logged effort is attested by the server; the chain does not verify learning.
- Mobile appA mobile web app brings the same Viva and course context to the phone, synced with the web.
A complete route through a sample semester. A real-world question still open.
What resonated with judges
After the demo, each judge said they wished they had something like Pathway as a student. One judge, speaking from their experience with ADHD, was especially enthusiastic about the nudges, accountability with study buddies, and the vision of connecting the tools students already use.
What shipped
Pathway became a working hackathon prototype: a named course, an inspectable reason, a modeled change, dated study blocks, and practice drawn from course material. The repo includes the repeatable sample profile, product captures, and model results.
What comes next
The model was trained and evaluated on synthetic data, not real students. Its boundary is not evidence that a study plan improves a grade. The next standard is student testing, clear observed-versus-assumed inputs, reliable cross-device plans, and an independent evaluation of academic outcomes.