Kunal
Aggarwal
role = "AI Model Analyst @ Deccan AI Experts"
focus = "LLM evaluation, RLHF, multi-agent systems"
obsession = "the almost-correct outputs. that's where it gets interesting."
➛ currently going deep on LangGraph + transformer internals
➛ exploring deep into AI, not tutorial-following, but the kind where you hit a wall at 2am.
01 / about
Who I Am.
I'm Kunal Aggarwal, a final-year CS undergrad at VIT Bhopal who builds and breaks language models for a living (well, for an internship). Currently working as an AI Model Analyst at Deccan AI Experts, where I dig into LLM outputs, run RLHF cycles, and chase down hallucinations like they owe me money.
I build AI things. Mostly I just try to understand why they break. My focus is on agentic AI systems, multi-step LLM pipelines, tool-calling agents, and retrieval-augmented generation. Real time with LangChain, LangGraph, smolagents, and Google ADK. I care about systems that actually work, not just notebooks that run clean.
Outside of AI, I build full-stack products with Flask and MySQL, spend too much time on LeetCode, and have opinions about prompt engineering I'll share if you ask.
Agentic AI systems, multi-agent frameworks, LLM evaluation, RAG pipelines. Building toward MCP (Model Context Protocol) projects to close the one gap in my stack.
India, Remote-first. Open to on-site opportunities in India or international remote roles in AI/ML.
LLM safety & alignment, RLHF systems, agentic reasoning, multi-modal AI, competitive programming.
02 / experience
Where I've Worked.
- Evaluated and debugged LLM-generated Python and SQL code across multiple model iterations, identifying logical hallucinations and syntax errors to ensure production-level correctness and consistency.
- Conducted RLHF (Reinforcement Learning from Human Feedback) by systematically ranking LLM outputs, contributing to an estimated 15%+ reduction in syntax errors and factual hallucinations across training cycles.
- Analysed complex technical prompts to surface edge cases in advanced algorithmic queries, improving model robustness and prompt-response alignment for Python and SQL code generation tasks.
03 / projects
Things I've Built.
Fine-Tuned Domain LLM
Fine-tuning Phi-3 Mini (3.8B) with QLoRA on the ChatDoctor-HealthCareMagic-100k dataset to build a domain-specialised medical question-answering model, trained end-to-end on a single RTX 4050. Built around a structured implementation plan with daily progress logs documenting the almost-correct outputs and what they reveal about the model.
github.com/koi-bito/Medical-QA-AI-AssistantAI-Powered Profile Generator
A LangChain parsing engine that maps natural language speech to structured JSON schemas via embeddings and generative AI, optimised for ATS keyword alignment. Modular async backend integrating OpenAI Whisper for transcription and multi-step LLM pipelines to convert raw audio into formatted, ATS-ready resume documents.
github.com/koi-bito/audio_to_resumeAI-Powered E-commerce Recommender
A full-stack e-commerce platform with a hybrid ML recommendation engine under the hood. Engineered TF-IDF + KNN with FULLTEXT search and cached REST API endpoints for personalised product recommendations at scale. Includes secure JWT auth, user behaviour tracking, and Chart.js analytics dashboards.
github.com/koi-bito/koimate04 / skills
My Toolkit.
05 / education
Where I Learned.
06 / achievements
Certifications.
featured
also completed
07 / contact
Let's Talk.
I'm actively looking for AI/ML internships where I can work on real LLM systems, agentic pipelines, or anything involving language models doing useful things in the world.
Got an interesting problem? Want to talk about RLHF? Or just want to connect, I'm easy to reach.
This opens your email client, no server needed.