AI engineering projects
Production Generative AI systems built at Dome Infosys, plus personal and academic work in retrieval, computer vision and OCR. Each production project names the part I engineered and who built the rest.
Production systems
- 01
Ziva
HealthcareA fertility assistant that answers from clinic-approved material over text and live voice — and refuses anything outside it.FastAPIQdrantGeminiOpenAILiveKitPostgreSQL - 02
Banau.ai
Home servicesNine AI capabilities behind one HTTP service, consumed by separate CRM, booking and customer products.FastAPIQdrantFAISSGeminiOpenAILangChain - 03
LEKHAN AI
PublishingTurns interviews, documents and photographs into a book — with a human verifying every fact before anything is drafted.FastAPICeleryQdrantGeminiPostgreSQLRedis - 04
Jobscater
RecruitmentJob posts generated from a few words, and a CV maker that scores like an ATS and explains what's wrong.PythonFastAPILLM APIsStructured JSON
Personal & academic work
Generative AI
KnowledgePilot
A question-answering API over your own documents, with citations and a small chat UI. It's where I worked out the thing I now put in everything: if the retrieved passages score below a threshold, say you don't know rather than generate from weak context.
- FastAPI
- LangChain
- Qdrant
- Streamlit
Computer Vision
Traffic sign detection for Nepal
My final-year project — a Faster R-CNN model recognising nine kinds of Nepalese traffic sign. No public dataset exists for these, so most of the work was collecting and annotating photographs rather than training.
- TensorFlow
- Keras
- OpenCV
Computer Vision
Nepali OCR
Reading Devanagari and English off scanned documents. EasyOCR does badly on a raw scan of Devanagari; nearly all the accuracy came from the OpenCV preprocessing in front of it.
- EasyOCR
- OpenCV
Generative AI
School chatbot
Notices and FAQs answered from a school's own documents, delivered as a widget they drop into their existing site with one script tag. Built for a client at Dome Infosys.
- FastAPI
- FAISS
- RAG