Project
Tayari Tutor Matching
Semantic tutor recommendation for the Tayari marketplace. I embedded tutor profiles with Ollama nomic-embed-text into LanceDB and search learner requests with filters for budget, subjects, and time-slot overlap, then ported it into the team’s Express backend as a tutor-matching service.
How it’s built View architectureHow it’s built
Tayari Tutor Matching · Semantic tutor recommendation using local embeddings and filtered vector search
Tayari Tutor Matching recommends tutors to language learners on the Tayari marketplace. I first trained a Keras scoring model on 300 synthetic student-tutor pairs, then rebuilt matching as semantic search: tutor profiles are embedded locally with Ollama nomic-embed-text in batches of 16 and stored in LanceDB, and learner requests are searched with SQL-style filters for budget, subjects and time-slot overlap. I then ported it into the team marketplace as an Express tutor-matching service.
- tutor profiles indexed
- 200
- student profiles in the test dataset
- 300
- tutors returned per request (top-k)
- 5
- input features in the earlier Keras ranking model
- 17
- test MAE of the Keras ranking model (45 test pairs)
- 0.031
- files I added to tayari-marketplace (9 commits, +1,767 lines)
- 27
Technology and what it does here
- Ollama nomic-embed-text
- Local embeddings for tutor profiles and learner queries via LangChain
- LanceDB
- Local vector table of tutors with filterable rate, slots and subjects
- LangChain (@langchain/ollama)
- OllamaEmbeddings batch and query embedding calls
- Next.js 16
- Prototype UI and /api/recommend, /api/get_tutors, /api/get_students routes
- Express (tutor-matching service)
- Matching, tutor, student and booking endpoints inside the marketplace
- TensorFlow / Keras
- Earlier 128-64-32 dense regressor scoring student-tutor pairs
- scikit-learn
- OneHotEncoder and StandardScaler for the Keras model features
- Papa Parse
- Cleans raw tutor and student CSVs into JSON
- Gemini (@google/generative-ai)
- API route prepared for match explanations; call currently commented out
Implementation details
- The marketplace version builds a LanceDB where clause with hard budget bounds, array_intersect on subjects and an OR group for time-slot overlap, then converts vector distance into a 0-100 score.
- Tutor indexing embeds concatenated profile text in batches of 16 and overwrites the LanceDB table in one write.
- The earlier approach engineered pair features (subject overlap, time-slot overlap, budget fit, session match, style compatibility) and trained a sigmoid-output dense network with early stopping.
System architecture
The semantic retrieval prototype encodes tutor and learner profiles in the same space. It filters for practical constraints when looking up similar tutors.
Technologies
- TypeScript
- LangChain
- Ollama
- LanceDB
- Express
- Vector search
