This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle
What I Built
I built LocalTutor AI, a fully offline, adaptive language learning application powered by local LLMs. It acts as a relentless but infinitely patient private tutor that lives entirely on your hard drive.
Traditional language apps force you down rigid, pre-baked skill trees (Iām looking at you, green owl). Human tutors are incredible, but they cost a fortune and demand strict scheduling. LocalTutor AI solves this by taking your target language, available daily time, and current skill level, and dynamically generating a customized curriculum day by day. It tests you, grades your answers, identifies your weak points, and seamlessly weaves those lagging topics into tomorrow's lesson.
Who would pay for this? Expats, corporate professionals, and serious self-studiers who want the personalization of a $50/hour private tutor for the price of a standard software subscription.
Demo
https://github.com/dinesh9-ai/Hacktoberfest2026_withAI/tree/main/DevChallenge3/Demo
Code
https://github.com/dinesh9-ai/Hacktoberfest2026_withAI/tree/main/DevChallenge3/Code
How I Built It
The core engine runs on Llama 3.1 (8B) served locally via Ollama, with a frontend built in Streamlit.
To make this actually work as a tutor rather than a stochastic parrot, I had to get highly specific with prompt engineering and state management. The app uses a local JSON database to persist user profiles, track daily streaks, and log historical performance.
The real magic happens in the daily generation loop. I enforce strict JSON schema outputs from Llama 3.1 to ensure the UI can cleanly parse native scripts (like Kanji and Hiragana), Romanization, and English translations into neat UI cards. The model takes the user's answers from the daily quiz, evaluates them against the day's lesson plan, and generates a structured "Report Card." Crucially, it extracts the exact concepts the user failed at and saves them to the local database as lagging_topics. The next day, the prompt dynamically injects those lagging topics into the system instructions, forcing the model to reteach the failed concepts before introducing new material.
Why Does Open Innovation Matter?
Building a dynamic, generative AI loop that runs multiple inference calls every day per user would be financial suicide if I were relying on a closed, pay-per-token cloud API.
Open-source AI made this side hustle viable for three critical reasons:
Zero Marginal Cost: Because Llama 3.1 runs locally on the user's machine via Ollama, my cloud API bill is exactly $0.00. I can charge a flat fee for the software without bleeding money every time a user decides to study for an extra hour.
Total Data Privacy: Language learners make embarrassing mistakes. Because everything runs offline, my users' data, typing habits, and learning struggles never touch a corporate server.
Unhinged Customization: I needed the model to strictly output native character sets and rigidly adhere to JSON schemas without cloud safety filters randomly deciding that a specific foreign idiom violated a content policy. Running open-weight models locally gives me absolute control over the inference parameters.
My First User
A close friend of mine recently landed a gig at a Japanese company. He loves the job, but he was constantly getting lost in translation during cross-team meetings. He decided he needed to learn Japanese and get his JLPT (Japanese-Language Proficiency Test) certification to survive the corporate battlefield.
He looked into local language institutes and quickly realized two things: the tuition fees for every progressive JLPT level were astronomical, and their rigid class schedules clashed horribly with his erratic software engineering hours.
I handed him the LocalTutor AI demo and told him to give it a spin. He used it consecutively for five days.
His feedback was exactly what you want to hear as a developer: he asked where he could enter his credit card. He told me the app completely eliminated the stress of the "institute schedule." He could study at his own pace, at 2 AM or during a lunch break. More importantly, he loved that the app adapted to his specific grasping speed. When he repeatedly failed to recognize specific Kanji characters, the app didn't drag him forward into a new chapter; it patiently looped those characters back into his next lesson until they stuck. He is now officially my first highly-motivated beta tester.
Prize Categories
- Open-Source AI
- Local Inference
- Best Side Hustle Concept
Top comments (0)