Shishir Adhikari
AI Engineer building production machine learning systems — Retrieval-Augmented Generation (RAG), Large Language Models (LLMs), computer vision, and NLP — deployed as scalable REST APIs with FastAPI, Flask, and Django.
I build the full pipeline: model development in PyTorch and Scikit-learn, vector search and semantic retrieval with FAISS, ChromaDB, and pgvector, LLM orchestration with LangChain and LangGraph, and production deployment with Docker. My work spans RAG assistants, Nepali/Devanagari OCR, and computer-vision face- and skin-analysis models — segmentation, detection with YOLO/RT-DETR, and ViT/DINOv2 classifiers. I also compete on Kaggle across computer-vision and NLP challenges to keep my modelling sharp.
Developed and deployed production REST APIs (FastAPI, Flask, Django) serving AI-powered products to real users. Built computer-vision face- and skin-analysis pipelines — face-shape detection plus skin models for acne severity (DINOv2), detection (YOLO/RT-DETR), wrinkle segmentation (U-Net++/EfficientNetV2), and age estimation (ViT). Engineered OCR systems including Nepali-language text extraction, a real-time facial emotion-recognition system (ViT + OpenCV, 8 classes), and an AI property-search system. Improved backend reliability with structured logging, error handling, and async optimization.
AI systems, APIs, automation
High-performance REST APIs
RAG, agents, LLM orchestration
FAISS, ChromaDB, Pinecone, pgvector
Model training & fine-tuning
Detection (YOLO/RT-DETR), segmentation, ViT
Fine-tuning, sentiment & text classification
Vision pipelines & landmarks
LLM integration & prompt engineering
Relational & document databases
Caching, sessions, real-time
Containerization & deployment
Cloud deployment, Workers, S3
Models, embeddings, transformers
ID: UC-2b2a53d7-1b67-46c7-bc74-00f917e9b045
View Certificate →Some things I've built recently.
Real-time computer-vision API classifying faces into 5 shapes at 85% accuracy (PyTorch + MediaPipe). Dual-LLM fallback (OpenAI GPT + Google Gemini) for 100% uptime, with Swagger docs, API-key auth, and gender detection driving hairstyle & beard recommendations.
RAG exam-question forecasting system (LangChain + FAISS) over 1,000+ pages on AWS S3 — 0.827 BERTScore F1, 89.3% syllabus precision across 44 held-out papers, sub-1.5s responses, and a 2.8% hallucination rate.
Multilingual RAG assistant over Nepali source material — Google Document AI OCR (43 pages → 950+ semantic chunks) with hybrid ChromaDB vector search + Devanagari keyword boosting. Recall@5 of 0.80, 4.10/5 answer correctness, 100% out-of-scope refusal; Streamlit + Django REST API on Hugging Face Spaces.
Production conversational-AI backend with Redis-backed session memory (<50ms retrieval) and ~1.2s end-to-end responses, built on FastAPI + LangChain with GPT for multi-turn, context-aware answers.
Scalable RAG pipeline using pgvector embeddings and FastAPI for low-latency semantic search and question answering over document collections.
Fine-tuned transformer models for Nepali sentiment analysis and hate-speech detection, tackling class imbalance in a low-resource language with focal loss. Built and trained end-to-end on Kaggle.
Have a question or want to work together?
