Agentic AI
Multi-agent workflows with tool calling, memory and human-in-the-loop checkpoints.
AI Engineer
Building trustworthy AI systems
> building AI agents
AI Engineer with 3 years of experience designing and delivering enterprise AI solutions: AI agents, RAG and workflow automation that improve how businesses work and decide.
About me
Professional summary
AI Engineer with 3 years of experience designing and delivering enterprise AI solutions that automate business processes and improve decision-making. I build AI agents, workflow automation, retrieval-augmented generation (RAG) and enterprise integrations using Python, FastAPI, LangGraph, LlamaIndex and modern LLMs, and I test them with golden datasets and LLM judges so they hold up in production.
I turn business requirements into scalable AI solutions and work closely with cross-functional stakeholders to drive AI adoption. I've delivered AI systems for teams in Dubai and India, and published NLP research at ICON 2024.
Ask about my work
Answers cite this page
Hi! Ask me anything about my projects, skills or experience. Every answer links to its source on this page, and I'll say so if I don't know.
Multi-agent workflows with tool calling, memory and human-in-the-loop checkpoints.
Hybrid search, reranking and cited answers over real documents.
Typed, read-only tools that plug any AI client into your data.
Golden sets, LLM judges, CI gates and prompt-injection defenses.
Technical toolkit
Career
Kinetics Group
May 2026 – Present · Dubai
Renie
From Jan 2025 · Dubai
ICFOSS
From Jun 2023 · India
KPMG
From Apr 2023
Featured work
Each project answers one question: how do you know the AI is right, and how do you stop it from being misused? Hover a demo to preview it, or click to watch.
RAG and evaluation
Cited answers on UAE VAT and corporate tax from seven official guides. It refuses questions the guides don't cover instead of guessing.
AI security
A prompt-injection game with five stacked defenses, backed by a benchmark that measures how often each defense breaks.
MCP server
Lets Claude Desktop, Cursor or any MCP client search the FastAPI docs and cite the exact section. Runs on a laptop CPU with no API keys.
AI reliability
An MCP server that checks an agent's answer against the tool outputs it actually received. It blocks wrong figures and invented claims, and rewrites the answer with citations.
Advertisement recommendation engine, a RAG chatbot on AWS Bedrock, LangChain review summarization and more.
Publications
NLP
An application that checks Malayalam scholarly writing for plagiarism against web content. Malayalam is a low-resource language with few tools of this kind, so the paper compares four text-matching algorithms to find the most reliable one.
Academics
Master's degree
Central University of Rajasthan
Bachelor's degree
University of Kerala
Get in touch
I reply fastest on email or WhatsApp, and I'm happy to walk you through any of these projects live.
MCP server
Recording: paraphrased questions find the right FastAPI docs sections with citation links, and a path-traversal attempt is refused.
An MCP server that lets Claude Desktop, Cursor or any MCP client search the FastAPI documentation and cite the exact section. It runs on a laptop CPU with no API keys, and the client's own model writes the answer.
search_docs, get_page, list_sections) with JSON-schema outputMeasured on 40 hand-written developer questions, every label checked by hand.
AI reliability
Recording: a draft client email about UAE VAT registration is checked against the FTA guides. Two wrong figures are blocked, and the corrected email cites the guide passages.
Agents call tools, get the right data back, and still misstate it in the answer. The Hallucination Firewall is an MCP server that records every tool output, splits the draft answer into single-fact claims, checks each one against that evidence, and returns a verdict per claim with the passage that proves or disproves it.
RAG and evaluation
Recording: two answers that cite the page they came from, then a refusal for a question the tax guides don't cover.
A question-answering system over seven public UAE Federal Tax Authority guides on VAT and corporate tax. Every answer cites its source page, and it refuses questions the guides don't answer instead of guessing.
AI security
Recording: all five levels, shown at 2× speed.
Talk an AI guard into leaking its secret password, then try again against five layers of defense. Behind the game is a reproducible attack benchmark that measures how often each defense breaks and how often it blocks harmless messages.