KANCA

AI-powered interactive storytelling that turns learning into a game kids actually want to play, grounded in Southeast Asian language and culture.

The Problem

Indonesia is in the middle of a literacy crisis that goes well beyond any one subject. PISA results show that over 60% of Indonesian students score at the lowest proficiency level across reading, math, and science, and a big part of why is simple: traditional learning materials are losing the attention war against Roblox, mobile games, and short-form video. Textbooks and storybooks ask for a child's focus; digital entertainment earns it.

We saw this up close in our own user research. One parent told us she wants her daughter to learn good values and life skills, but every book or lecture she tries puts the kid to sleep. Her daughter was blunter about it: the stories Mom gives her are boring, she'd rather just play Roblox. That's the gap KANCA is built to close, not "kids don't have access to good educational content," but "the format that content comes in can't compete for a child's attention."

Because the team built KANCA's first full deployment for the BI-OJK Hackathon 2025, whose theme was financial literacy, the concrete proof point we shipped was a financial-literacy story engine, complete with OJK-aligned content and a parent/teacher dashboard. But the underlying problem, and the underlying engine, isn't finance-specific. The same architecture is built to teach any subject or value through the same loop: let the child build the story, not just read it.

The Solution — "Belajar Ga Harus Membosankan!" ("Learning Doesn't Have to Be Boring")

KANCA turns a lesson into a story the child creates and plays through, rather than one they're handed. The flow is the same regardless of subject:

Build the story. The child picks a theme (school, fantasy, shopping, adventure), a character, and the values or concepts to weave in.
Play it out. The story plays with illustrated visuals and voiceover, no passive reading.
Make the call. At key decision points, the child chooses how the character acts. Good decisions earn rewards and praise in the moment; bad decisions play out their consequences and get gentle, constructive feedback.
Land the lesson. The story closes with a plain-language summary of the core value it taught and concrete examples of how to apply it day to day.

Daily quizzes, XP, levels, and badges keep the loop going so a child comes back the next day on their own, instead of needing to be told to.

For the BI-OJK Hackathon specifically, that engine was pointed at financial literacy: stories about saving, sharing, and honest dealing, grounded in OJK's official curriculum. The same engine, pointed at a different content corpus, teaches a different subject entirely.

The LLM Layer: Model-Agnostic, Currently Running on SEA-LION

The retrieval-augmented generation pipeline underneath KANCA isn't tied to any one model. Depending on the deployment, it can sit on top of global providers (the BI-OJK build initially used OpenAI and Azure, with a roadmap toward local providers like SahabatAI) or a Southeast Asia-tuned model.

For KANCA's entry into AI Singapore's SEA-LION competition, the model in the loop is SEA-LION itself, AI Singapore's own LLM, fine-tuned specifically for ASEAN languages and cultural context. That wasn't just a competition requirement; it's a genuinely better fit for the content KANCA generates. A generic Western-tuned model tends to produce stories that read like a translated fable. SEA-LION's tuning toward Bahasa Indonesia and regional cultural references makes the difference between a story that's merely correct and one that actually feels native to an Indonesian child.

Story generation is grounded with RAG rather than left to free-form generation, regardless of which model sits underneath. We index the relevant subject-matter literature (for the financial-literacy build: OJK's official guide for teaching financial concepts by age group, Kemendikbud's collection of moral-value folktales tied to the Pelajar Pancasila student profile, OJK's "Mengenal Jasa Keuangan" series, and the official kindergarten/elementary student and teacher textbooks) into a vector database. When a child requests a story, the system retrieves the most relevant passages and feeds them to the LLM as context, which writes a branching narrative with built-in decision points, checked against the source guidance rather than invented from scratch. The generated story then goes to an image model (Google's Gemini image-generation model, internally nicknamed "Nano Banana") for scene illustrations and to Chirp 3 for text-to-speech narration, producing a fully illustrated, voiced story rather than plain text.

Personalized Insights for Parents and Teachers

The same retrieval pattern powers the side of KANCA adults see. A second RAG pipeline pulls in a child's own gameplay history (which decisions they made, how consistently, on which concepts) alongside the same curriculum guidance used for story generation, and a retrieval agent first works out what the parent or teacher is actually asking before pulling only the relevant slice of that child's data. The LLM synthesizes both into a personalized, curriculum-aligned summary, surfaced either as a dashboard (playtime, stories completed, success rate, and concept-level trends over a week or month) or as an answer from the in-app chatbot, available to parents as a virtual coach around the clock.

Architecture & Engineering

KANCA ships as a single cross-platform codebase covering Android, web, and iOS, with cloud deployment and a CI/CD pipeline for continuous delivery. The four-person team split cleanly across AI/ML and the RAG pipelines, mobile development, UI/UX design built around a child's attention span rather than an adult dashboard, and backend/DevOps for the data layer and deployment.

Gameplay data isn't just logged for analytics, it's the substrate the insight engine reasons over. Every decision a child makes inside a story becomes a data point for measuring concept understanding, since comprehension is inferred from the share of correct in-story decisions tied to a given concept, tracked separately from raw engagement metrics like time played or stories completed. Keeping those signals distinct is what lets the dashboard say something more useful than "your kid played for 30 minutes today."

Pilot Results

We piloted KANCA at a primary school in Semarang City before treating any of this as a hypothesis. 92% of students said they preferred KANCA's stories over their regular storybooks, 88% said being able to choose their own story made learning more enjoyable, and the app scored 4.6 out of 5 on a retention-and-replay measure. On the adult side, teachers and parents reported roughly 40% more time freed up for child development work instead of manual evaluation, since the dashboard handles that synthesis automatically.

Business Model & Market

KANCA's go-to-market, scoped around the financial-literacy vertical built for the BI-OJK Hackathon, runs a dual-sided model: a B2B SaaS license sold to schools, tiered by enrollment size, where one teacher-generated story can be played by an entire class while each student still gets an individual insight profile, and a B2C subscription for parents, ranging from a free tier (five free stories plus one daily random story) up to a Premium+ tier with monthly insight reports and chatbot access.

Indonesia has roughly 33 million kindergarten and elementary-age children, putting the addressable market north of Rp 13 trillion a year at current per-child pricing assumptions. The initial target is the roughly 4 million children in technology-ready private and SPK (Sekolah Penyelenggara Kerjasama) schools, with a three-year goal of 60,000 children across 85 partner schools and 8,000 B2C subscribers. Because the underlying engine isn't tied to a single subject, this addressable market grows further as additional content verticals beyond financial literacy come online.

Future Work

The pilot validated the core loop, kids genuinely prefer this to a book, but the next phase is about depth and reach rather than novelty. That means scaling toward roughly 1,500 schools nationwide, expanding the content engine beyond financial literacy into other subjects and values, and, since the underlying RAG architecture isn't tied to Indonesia or to any single model, extending the same offline-friendly, story-first pattern into the broader Southeast Asian market. On the policy side, KANCA lines up with Kemendikbud's plan to introduce AI and coding as elective subjects starting in the 2025/2026 school year, giving both students and teachers a lower-stakes way to get comfortable with AI-assisted learning before that rollout goes wide.

Why the Name "KANCA"?

"Kanca" comes from the Javanese word konco, meaning "friend." The name is the mission statement: not a textbook with a chatbot bolted on, but a learning companion a child actually wants to spend time with.

Built by KancaKids

KANCA was built by Team KancaKids, four students from BINUS University's Semarang campus: James Alvin Dhanardi (AI/ML, RAG systems), Wisnu Putra (Backend & DevOps), Hibatullah Fawwaz Hana (Mobile Development), and Nicholas Lorenzo Setiawan (UI/UX Design). The financial-literacy build of KANCA was developed for the BI-OJK Hackathon 2025 (under FEKDI x IFSE); a parallel build was entered into AI Singapore's SEA-LION competition.

Sources

Badan Pusat Statistik & Otoritas Jasa Keuangan. (2025). Survei Nasional Literasi dan Inklusi Keuangan (SNLIK) Tahun 2025.
OECD. (2023). OECD/INFE 2023 International Survey of Adult Financial Literacy (OECD Business and Finance Policy Papers No. 39). OECD Publishing. https://doi.org/10.1787/56003a32-en
IMARC Group. (2024). Indonesia Edtech Market Size, Share, Trends and Forecast by Sector, Type, Deployment Mode, End User, and Region, 2025–2033.
AI Singapore. (2026, January 6). Kanca [Case study]. SEA-LION. https://sea-lion.ai/case-study/kanca/