Azerbaijani NLP Tutor AIC
Budget: $15 – $25 USD
I’m building YapayAITeach, a learning companion designed to slot straight into the Azerbaijani school system, and I now need an NLP-focused developer to push the core intelligence forward. The current prototype handles lesson sequencing but lacks a robust language layer that truly understands and responds in Azerbaycanca, so your mission is to craft that capability from the ground up.
Here’s what I have in mind:
• Train and fine-tune a natural-language model that performs conversational Q&A, short-answer grading and topic summarisation entirely in Azerbaijani.
• Implement intent detection plus basic sentiment to adapt the tutor’s explanations to each student’s mood and performance.
• Package the model behind a simple REST or GraphQL API so my existing front-end (React) can call it with minimal latency.
• Provide evaluation scripts and small benchmark datasets so we can track accuracy over time.
I’ll share the current codebase (Python, FastAPI) and an initial corpus of curriculum texts. Feel free to lean on Hugging Face Transformers, spaCy, or any toolkit you prefer, as long as the final model runs on commodity GPUs.
Acceptance criteria
1. At least 85 % accuracy on the supplied comprehension-question set.
2. Average response time under 800 ms on a T4 GPU.
3. Clear deployment guide and a one-click Docker compose file.
If you’ve shipped production NLP in low-resource languages—or specifically Azerbaijani—let’s talk. I’m ready to integrate your work as soon as it lands in the repo.
Here’s what I have in mind:
• Train and fine-tune a natural-language model that performs conversational Q&A, short-answer grading and topic summarisation entirely in Azerbaijani.
• Implement intent detection plus basic sentiment to adapt the tutor’s explanations to each student’s mood and performance.
• Package the model behind a simple REST or GraphQL API so my existing front-end (React) can call it with minimal latency.
• Provide evaluation scripts and small benchmark datasets so we can track accuracy over time.
I’ll share the current codebase (Python, FastAPI) and an initial corpus of curriculum texts. Feel free to lean on Hugging Face Transformers, spaCy, or any toolkit you prefer, as long as the final model runs on commodity GPUs.
Acceptance criteria
1. At least 85 % accuracy on the supplied comprehension-question set.
2. Average response time under 800 ms on a T4 GPU.
3. Clear deployment guide and a one-click Docker compose file.
If you’ve shipped production NLP in low-resource languages—or specifically Azerbaijani—let’s talk. I’m ready to integrate your work as soon as it lands in the repo.