विपुल!

Inclusive Hackathon Learnings & Success Story

*Ferizaj, Kosovo (Central Hotel) · 10–12 July 2026 · *

The first ADT inclusive hackathon ran over three days, Friday to Sunday, 10–12 July 2026, at the Central Hotel in Ferizaj. It brought together ~35 people in total: four competing teams of University of Pristina students, embedded teachers, technical mentors, judges, and organisers. The event met its two goals at once — it produced solutions teachers would genuinely use, and it validated a methodology we can now re-run in other countries.

The single most important design choice, and the clearest success signal, was deliberately keeping the event small. With only six teams (initially planned, then made 4) and a tight headcount, mentors and teachers could sit close to each group, guide continuously, and give real attention rather than triage. Every team reached a working, demoable solution.

The headline learning Fewer participants, better results. A leaner event let us engage and guide each team closely, and all four teams shipped something demoable. For a first event built to be replicated, close guidance beat scale.

The challenge

The overarching question: How can AI support and strengthen the work of teachers and teacher-assistants in inclusive classrooms?

Teams chose from four streams, each grounded in a gap teachers named directly. The test for every stream was the same: would a teacher in the room actually use it on Monday?

Stream Question Grounded in
1. Content & Materials How can AI transform classroom materials to work for every learner? Teachers said content isn't accessible and appropriate materials are scarce.
2. Assistive Tech & AAC How can AI help every teacher discover, learn, and confidently use the right AT for each student? The barrier isn't the technology; it's knowing it exists and how to use it.
3. Assessment & Outcomes How can AI help teachers connect what a child does today to where the curriculum wants them to go? Internal assessment of students with disabilities and cross-subject planning are daily struggles.
4. Bring Your Own Challenge What gap do you see that we don't? Feedback boards named stigma, large classes, training, weak school-company collaboration, and more.

Results & awards

All four teams reached a working, demoable solution — a 100% completion rate against our own "60%+ demoable" target. Awards followed the four judging criteria, so recognition mapped directly to what we told teams mattered.

Category awards

Four individual/team awards, one per judging criterion:

Award (criterion) Points Recognises Awarded to
Most real-world use case (Useful to the real problem) / 35 Hits a gap teachers actually named; usable next week. Team 2 — MësoLehtë AI
Most accessible (Accessibility) / 30 Verifiably works with assistive tech; craft is evident. Team 4 — Mësim i Qartë
Most complete (Actually works) / 25 Functional end to end; hand it to a user now. Team 3 — AtlasPlan
Most impactful (adapted from Demo clarity) / 10 Clear problem, solution, and impact; strongest overall reach. Team 1 — Teacher Workspace

Winner — Team 2, MësoLehtë AI

A teacher uploads a difficult text and AI adapts it to each student's needs — simplified levels, vocabulary, quizzes, TTS in Albanian and English — while the teacher reviews, edits, and publishes. AI proposes; the teacher decides.

Runner-up — Team 3, AtlasPlan

A privacy-first planning workspace that anonymises student data, analyses PIA plans, and builds individualised schedules and parent reports, all fronted by a warm, accessible interface.

The other two builds

Judges

How we planned and ran it

The event followed the Running an Inclusive Hackathon guidebook: define the purpose, recruit through existing networks, run virtual build-up workshops, then a fully in-person build. Recruitment ran through the University of Pristina and partner networks. Three virtual build-up sessions (the problem; tools, scope and teams; open-source contribution) formed teams and closed skill gaps before Day 1, so in-person time went to building rather than orientation.

People in the room. Every team had an embedded teacher as a lived-experience guide and technical mentors from Axians (Arbnor Bajrami, Blerta Jashari, Enrik Dodaj) to unblock the build. Organising and judging was led by Elias Constantopedos, Timur Ramiqi, Laurat Raca, and Vipul Siddharth across UNICEF Office of Innovation and UNICEF Kosovo.

Accessibility, operationalised. Provisions weren't aspirational — they ran. Silent corners and an accessible restroom sat at the hacking area, and because participants were accommodated in the same hotel, they could retreat to their own rooms to recharge beyond the quiet spaces on the floor. Interpretation, mediation, and specialist support were staffed against the confirmed participant mix.

Learnings for the next event

These are the operational lessons that only became visible on the day. They are the most reusable part of this report — the spine of the methodology we hand to the next country.

1. Accessibility has to be checked everywhere, including the rooms you don't run

We got the hacking area right — silent spaces, an accessible restroom on hand. What we missed was checking every room where participants were accommodated; we never actually measured the toilet and entrance door widths of the accommodation rooms. That surfaced on the day and forced a last-minute room swap, which thankfully worked. The lesson: the accessibility walk-through has to cover the whole participant journey — sleep, eat, retreat — not just the build floor. Measure door widths, don't eyeball them.

2. Send written tool setup before the event, not just workshops

There will always be cases of tools not being ready — Codex, Git, the GitHub CLI. We ran build-up workshops on exactly this, and they helped, but they weren't enough on their own. Next time, ship written step-by-step setup instructions before the event so participants arrive with working environments and the workshop reinforces rather than installs.

3. Teach best practice relentlessly, especially in the age of AI

We reminded teams to commit and push, but not often enough. This matters more now, not less. A single AI prompt can take a solution in a direction there's no returning from — unless the team has been practising version control diligently. When generation is fast and cheap, the discipline that protects work is the one thing AI won't do for you. Frequent, visible reminders to commit and push are a safety mechanism, not a nicety. The fundamentals — version control, clean handovers, knowing what your tools are actually doing — become more valuable as the generation gets easier, because they're the difference between a recoverable project and a lost afternoon. Teaching the plumbing before the shortcut isn't nostalgia; it's what keeps AI-assisted work salvageable.

4. Have a feedback form ready for the moment the event ends

We planned to share a feedback form with certificates as the hook — but we should have had one ready to hand out right after the event, while people were still in the room. All our feedback so far is verbal. Anonymous, written feedback would add real value: people say things on a form they won't say to your face.

5. Brief judges earlier — well before demo day

We briefed judges on the challenge and rubric before the demos, but that was too late. One judge was visibly confused during scoring, which leaves a bad impression and risks the fairness of the result. Send judges the streams and the scoresheet days ahead, and treat the day-of briefing as a refresher, not a first read.

6. Run a strict, visible timer during demos

The first demo overran, and we absorbed it — which then forced us to extend everyone else's slot to keep the opportunity equal. That's a fairness problem dressed up as generosity. Set a strict, visible countdown for both presentation and Q&A, enforced from the first team, so equal time is built in rather than retrofitted. It also protects the interpretation budget, which stretches every timed segment.

What worked, and worth keeping