I build transparent, scalable AI systems focused on efficiency and governance, specializing in model optimization (distillation & quantization), domain-adapted tokenization, and knowledge-graph–powered reasoning.
My work includes training custom tokenizers, resizing model embeddings to align with bespoke vocabularies, and fine-tuning lightweight LLMs for specialized domains.
Across multiple projects, I’ve delivered LLM systems with grounding, explainability, observability, and deployment controls, applying consistent safety- and governance-first system design principles across search, conversational AI, and decision-support systems.
- Core | APIs & Deployment
|
- MLOps & AI Tools | Data & Tools
|
-
Domain-Adapted Tokenizer & TinyLLaMA v1 QLoRA Fine-Tuning – Tokenizer Training, Embedding Resize & QLoRA Fine-Tuning:
Trained a BPE tokenizer on a 500 M-token biomedical corpus; resized TinyLLaMA v1’s embedding matrix to fit the new vocabulary; applied 4-bit QLoRA with PEFT adapters; fine-tuned for domain QA—achieving a 12% drop in perplexity and a 15% F1 lift.
Tools: Python, Hugging Face Transformers, Tokenizers, BitsAndBytes, PEFT, PyTorch -
Model Optimization Pipeline – Distillation, Quantization & Deployment:
Trained and distilled a CNN for digit recognition; exported to TorchScript and achieved >98% accuracy post-quantization; built a FastAPI backend and Streamlit UI for real-time inference with Supabase logging; containerized with Docker and deployed on GCP Cloud Run; tracked performance with MLflow and W&B.
Tools: TorchScript, Distillation, Quantization, FastAPI, Streamlit, Supabase, Docker, GCP, MLflow, Weights & Biases -
Conversational AI with Knowledge Graph:
Built a chatbot combining GPT, LangChain RAG, and Neo4j to deliver grounded, context-aware responses, improving reliability and traceability in conversational systems.
Tools: Python, GPT, LangChain, RAG, Neo4j -
Neural Search:
Developed a neural search engine using SentenceTransformer embeddings and a vector database to deliver high-accuracy semantic retrieval.
Tools: Python, NLP, SentenceTransformer, Vector Database -
Explainable AI (XAI) for Fraud & Decision Support:
Developed explainability workflows using LIME, SHAP, and DiCE to support investigation, audit, and accountable decision-making; demoed an interactive XAI prototype.
Tools: Python, LIME, SHAP, DiCE
- Neo4j Certified Professional & Neo4j Graph Data Science — Neo4j
- Generative AI Engineering.
- Applied Data Science Lab — WorldQuant University
- Red Teaming LLM Applications — Giskard AI & DL
- Quality & Safety for LLM Applications — WhyLabs & DL
- OpenTelemetry Observability Lab — Linux Foundation
- Explainable AI: Developed tools with LIME, SHAP, and DiCE to enhance AI transparency; demoed an XAI prototype on Hugging Face Spaces.
- Sovereign AI Governance — Independent work exploring policy-aligned, auditable, and safety-aware AI system orchestration.