MarketMatic
A year of research into AI-driven marketing tooling, ending in a peer-reviewed IEEE paper.
Small businesses need dependable customer support and marketing copy, and can rarely staff either. MarketMatic does both: a retrieval-augmented assistant that answers only from a tenant's own uploaded documents, and a generator that writes marketing copy in English and Sinhala.
Final-year research at SLIIT.
- ROLE
- Researcher & Developer
- YEAR
- 2025–2026
- TEAM
- SLIIT · Supervised by Dr. Kapila Dissanayaka and Ms. Fathima Fanoon
MarketMatic: Multi-Tenant Customer Support Smart Assistant with RAG for Small and Medium Enterprises
Sandipa U K S; Rajapakshe R.S.L.; Jayasuriya J.A.D.C.G.; Wijayarathna H.P.I.G.; Kapila Dissanayaka; Fathima Fanoon
- PUBLISHED IN
- 2026 18th International Conference on Electronics, Computers and Artificial Intelligence (ECAI)
- DATES
- Conference 02–03 July 2026
RETRIEVAL-AUGMENTED SUPPORT
Each tenant's documents sit in an isolated vector collection. Answers come from that business's own material, and cite it.
BILINGUAL GENERATION
GPT-2 Medium and mT5-base, LoRA-adapted on a 25,000-example marketing dataset we built ourselves.
SINHALA THAT RENDERS
A HarfBuzz shaping stage fixes the ligatures and diacritics that garble Sinhala in most pipelines.
DEGRADES RATHER THAN STOPS
Every stage has a defined fallback, so a rate-limited dependency slows the service instead of ending it.
- BACKEND
- Python · FastAPI
- ML / NLP
- PyTorch · Hugging Face · GPT-2 Medium + LoRA · mT5-base + LoRA · 4-bit quantisation
- RETRIEVAL
- ChromaDB · Vector search · Cross-encoder reranking
- RENDERING
- HarfBuzz (uharfbuzz) · Playwright + Chromium · Pillow
- DATA
- Custom 25,000+ example bilingual dataset (EN / SI)
_