AI Portfolio

LLM Fine-Tuning, RAG and Agentic AI Projects

A look at AI builds: a custom fine-tuned and quantized LLM, a retrieval-grounded knowledge assistant, an agentic tool-calling system, and full-stack systems work.

Qwen 2.5-3B Ayurveda Expert DevDocs RAG Assistant Agent Console Library Management System
AI Project • LLM Fine-Tuning

Qwen 2.5-3B Ayurveda Expert

Custom LLM Fine-Tuning, Quantization & Deployment. Domain: Ayurveda (Traditional Medicine).

A domain-specialized language model built by fine-tuning Qwen2.5-3B with LoRA on Ayurveda subject matter, then quantizing and packaging the result for deployment. The goal was a compact model that answers Ayurveda-specific questions accurately without the cost or latency of routing every request through a general-purpose frontier model.

Google Colab notebook showing the Qwen 2.5-3B LoRA fine-tuning run for the Ayurveda expert model, with training progress and the saved checkpoint

Fine-Tuning Approach

Used LoRA (Low-Rank Adaptation) to fine-tune Qwen2.5-3B on Ayurveda domain data with a Hugging Face Trainer and DataCollatorForLanguageModeling, keeping the base model frozen and training only lightweight adapter weights, so the specialization stays fast and cheap to iterate on.

Training Setup

Mixed-precision (fp16) training on Google Colab, with a held-out evaluation split tracked alongside the training set so generalization could be monitored as the run progressed, not just training loss.

Quantization & Deployment

The fine-tuned model was quantized for efficient inference and packaged for deployment, trading a small amount of precision for a significantly lower memory footprint and faster response times.

Result

A saved, deployable checkpoint (qwen2_5-3B-ayurveda-expert-final) specialized for Ayurveda question-answering, ready to serve without depending on a general-purpose LLM API for every query.

AI Project • Retrieval-Augmented Generation

DevDocs Assistant

RAG-Based AI Knowledge Assistant. Internal Engineering Documentation.

An internal knowledge assistant that answers questions about engineering docs, APIs and infrastructure using only indexed internal documentation, so answers stay grounded in what's actually written instead of drifting into plausible-sounding guesses.

DevDocs Assistant chat interface answering a grounded question about API rate limits, with citations and a retrieval trace panel showing the FAISS pipeline and retrieved chunks

Retrieval Pipeline

Documents are ingested, chunked, embedded and indexed in FAISS. Each query is embedded, matched against the index by cosine similarity, reranked, and only the top-matching chunks are passed to the model to generate an answer.

Grounding & Citations

Every answer is checked against its retrieved source chunks before being shown, tagged "Grounded" with a citation count, and flagged if it contains claims the sources don't support, instead of letting the model answer from general knowledge.

Source Transparency

A retrieval trace panel shows exactly which document chunks were retrieved and their similarity scores, plus a 2D projection of the embedding space for that query, so anyone using the assistant can check an answer against the underlying doc rather than taking it on faith.

Performance & Evaluation

The showcase build's own metrics panel reports a 42ms average retrieval latency, 1.8s average end-to-end latency, 94.6% answer groundedness and a 1.8% hallucination rate on its evaluation set, tracked to catch regressions as the retrieval pipeline changes.

Demo Scale

The showcase build indexes 4,218 internal documentation pages into 38.9K chunks, enough to demonstrate the ingest → embed → search → rerank → generate pipeline behaving correctly at a realistic scale.

DevDocs Assistant evaluation metrics panel showing retrieval latency, end-to-end latency, answer groundedness, hallucination rate and a 2D embedding-space projection
AI Project • Agentic AI

Agent Console

AI Agent & Agentic Workflow System. Multi-Step Reasoning & Tool Calling.

A demonstration console for agentic AI workflows: LLMs that plan across multiple steps, choose the right tool for each step, call external APIs and retrieval systems, and hand back a finished result instead of a single reply. Built around tool calling, orchestration and prompt design with LangChain and Python.

Reasoning Loop

Each task runs a thought → action → observation loop: the agent reasons about what it needs, takes an action (like a web search or API call), observes the result, and decides the next step, live-logged in a running trace console.

Tool Calling & Orchestration

Built with LangChain in Python, the agent selects from a registry of tools for each step of a task rather than following a fixed script, so it can handle tasks that need a different tool path depending on what it finds along the way.

Observability

A live trace panel logs every thought, action and observation as a task runs, and an architecture/metrics view tracks tool-call success rate and steps per task, so agent behavior stays inspectable instead of a black box.

Demo Scale

The showcase console's own metrics panel reports several thousand completed demo tasks with a high tool-call success rate, used to validate the orchestration logic across varied task types.

Full-Stack Project • Systems

Library Management System

Full-Stack Systems Build. Library Catalog, Members & Loans.

A full-stack library management system covering catalog, membership and loan tracking across multiple branches, with a Go backend exposing a REST API and an admin console for day-to-day staff use.

Backend

Built with Go using the Gin router, exposing REST endpoints for books, users and loans (GET/POST/DELETE), returning structured JSON for catalog and availability data.

Catalog & Availability

Tracks call numbers, ISBNs, copy counts and live availability status (available vs. on loan) per title, across all branches from a single admin view.

Admin Console

A staff-facing dashboard for browsing the catalog and checking loan status, with a built-in API console for testing endpoints directly during development.

Data Model

Structured around books, members and loans as first-class records, with fields like copies-available and created-at timestamps returned directly from the API.

Have an AI Idea in Mind?

Tell us about the workflow, chatbot or automation you're picturing, and we'll help you scope it.

Start a Project WhatsApp