AI Systems & Machine Learning
Designing, training, and deploying end-to-end intelligent systems — from LLM fine-tuning and RAG pipelines to multimodal CLIP search and graph representation learning.
Core Focus Areas
LLM Engineering
Prompt engineering, instruction tuning, evaluation frameworks, and building production LLM pipelines with structured output and tool use.
RAG Systems
Retrieval-augmented generation: document ingestion, chunking strategies, dense retrieval with sentence-transformers, hybrid search, and grounded generation with citations.
Fine-Tuning
End-to-end fine-tuning pipeline for Qwen2.5 LLMs: dataset engineering, SFT/LoRA/QLoRA training via mentorai-cli, evaluation (ROUGE/BLEU/BERTScore/LLM-as-Judge), and deployment as the model backing MentorAI.
Multimodal AI
End-to-end multimodal product search engine: CLIP (ViT-B/32) embeddings + FAISS IndexFlatIP across 145,050 products, with zero-shot tagging, weighted fusion (60/40 image/text), offline recommendations, and Kafka-tracked search events feeding Airflow-scheduled daily reindex.
Machine Learning
Supervised and unsupervised ML for classification, regression, risk scoring, anomaly detection, graph-based learning, and reinforcement learning. Production deployment with MLOps practices.
AI System Pipelines
Concrete architectural patterns used across production AI and ML projects.
RAG Pipeline (MentorAI)
Documents
PDF, Markdown, Code
Chunking
512 tokens + overlap
Embeddings
BAAI/bge-small-en-v1.5
Vector Store
FAISS IndexFlatIP
Retrieval
Top-K dense search
LLM + Context
Qwen2.5 (QLoRA fine-tuned)
Grounded Response
With citations
MentorAI Fine-Tuning Pipeline (LLM Fine-Tuning Framework)
Knowledge Units
Domain source material
Dataset Engineering
Alpaca / ChatML / ShareGPT formats
Tokenization
Reusable pipeline
Training
SFT → LoRA → QLoRA (4-bit)
Evaluation
ROUGE, BLEU, BERTScore, LLM-as-Judge
Model Export
Adapter merge, Safetensors
Inference API
FastAPI, feeds MentorAI
VisualMind Search Pipeline (VisualMind)
Query
Image, text, or both
CLIP Encoder
ViT-B/32, 512D embedding
Multimodal Fusion
Weighted average (60/40 image/text)
FAISS IndexFlatIP
145,050 vectors, exact search
Results + Tags
Zero-shot category/color/style
Kafka Events
Search telemetry
Airflow Rebuild
Daily index + recommendation refresh
Featured AI Projects
VisualMind
Traditional text-based product search misses visual similarity — users can't find what they're looking for with words alone.
MentorAI
Technical learners waste hours searching docs and tutorials when they need a specific, contextual answer grounded in their actual study material.
MentorAI Fine-Tuning Pipeline
No reusable, production-ready pipeline exists for fine-tuning open-source LLMs on domain-specific data — every project starts from scratch.
Email EU Graph Platform
Organisational communication patterns are invisible — email data contains rich relational structure that standard tools cannot exploit.
Plant Biology Statistical Analytics Pipeline
Biological research experiments feature complex multivariate interactions across treatments and genotypes, requiring rigorous, reproducible statistical modeling to extract publication-grade insights.
Smart Email Assistant
Repetitive email work — classification, prioritisation, drafting standard replies — consumes hours that could be spent on meaningful work.
Sentinel-AI
Financial risk analysis requires integrating heterogeneous data sources — market OHLCV data, news sentiment, chart patterns, and regulatory filings — into a unified, explainable risk scoring system.
AI & ML Toolchain
Python
Core ML & pipeline development
PyTorch
Deep learning & fine-tuning
HuggingFace
Transformers, PEFT, TRL
FAISS
Dense vector retrieval
LangChain
RAG pipelines & LLM orchestration
MLflow
Experiment tracking & model registry
Scikit-learn
Classical ML & feature engineering
Pandas
Data manipulation & analysis
NumPy
High-performance vector operations
Kafka
Event streaming pipeline
Airflow
Workflow orchestration & scheduling
Docker
Containerized deployment
06 Certified AI/ML Specializations
Stanford University, DeepLearning.AI, and DataCamp verified credentials.