Available for opportunities

Hi, I'm
Amrutha Thalla

Generative AI & Agentic AI builder — RAG, multi-agent systems, fine-tuning, and applied comparisons across frameworks — with a data analyst foundation underneath.

9
GenAI / ML projects
RAG · Agents
+ Fine-Tuning
M.S.
Indiana Tech, USA

Who I am

I build with Generative AI — retrieval-augmented generation, agentic systems, and parameter-efficient fine-tuning. My background is in Information Systems (M.S., Indiana Institute of Technology), which gave me an entry point into NLP through the field's older questions of how knowledge is organized and made findable, rather than through model architecture first. I learn by building, comparing, and reporting what actually happened — several of my projects deliberately implement the same task two ways, specifically to surface real differences rather than assume one approach represents a whole technique. Alongside GenAI work, I have a solid foundation in data analysis, Python, SQL, and machine learning.

Academic background

IN PROGRESS

Full Stack Data Science with GenAI & Agentic AI

Naresh i Technologies

Hands-on training in data analysis, machine learning, LLMs, LangChain/LangGraph, RAG, fine-tuning, and agentic AI application development.

Master of Science — Information Systems

Indiana Institute of Technology, Fort Wayne, USA

Aug 2023 – May 2025

Focus areas: Data Management, Cloud Computing, System Architecture.

B.Tech — Computer Science & Engineering

JNTUH, India

2018

Foundation in algorithms, databases, object-oriented programming, and software engineering principles.

Technical toolkit

GenAI & Agentic AI
RAG LangChain / LangGraph CrewAI phidata LoRA Fine-Tuning Hugging Face Vector Databases Prompt Engineering
ML & Data Science
Machine Learning Deep Learning Scikit-learn Pandas / NumPy
Programming
Python SQL C# / .NET Core JavaScript
Cloud & Infra
Azure Docker Git REST APIs

Featured projects

Five projects below, each demonstrating a genuinely distinct GenAI mechanism. Several are deliberate comparisons — two frameworks, two QA approaches — built to surface real findings rather than a single success story.

⚡ Featured

RAG Over RAG Research Papers

GenAI Completed

A retrieval-augmented Q&A system answering questions about six foundational RAG papers, grounded in the papers' own text rather than the model's general knowledge — using the method to study the method.

Python LangChain sentence-transformers ChromaDB Groq API (Llama 3.1) Streamlit
✂️
Chunk & Embed
500-word chunks, 50-word overlap
🔎
Retrieve
Top-k semantic similarity search
💬
Generate
Grounded answer + source citations
🛑
Guardrail
Flags insufficient evidence instead of guessing
Design decisions
Chunk size/overlap directly affected retrieval accuracy
Explicit "don't guess" instruction reduces hallucination
Tested with
Out-of-scope questions, confirmed correct refusal

Multi-Agent System: CrewAI vs. phidata

GenAI Completed

The same two-agent research-and-summarize task built in two different agent frameworks, compared side by side in a Streamlit app — to isolate whether framework choice or tool choice actually drives the difference in output, rather than assuming one implementation represents "agent frameworks" generally.

Python CrewAI phidata Gemini API Serper / DuckDuckGo Streamlit
🔍
Research Agent
Searches the web for current findings
✍️
Writer Agent
No search tool — works only from given findings
⚖️
Side-by-Side
Same topic, both frameworks, compared live
Real test result
CrewAI + Serper: found real info, conflated two dated events, flagged its own uncertainty correctly
 
phidata + DuckDuckGo: found nothing, explicitly said so rather than guessing

LangChain Agent over a Pandas DataFrame

GenAI Completed

An agent that decides which Pandas operations to run to answer natural-language questions about a retail sales dataset, tested specifically on multi-step questions — filter, then group, then aggregate, including a nested per-group "top item in each category" query.

Python LangChain Groq API (Llama 3.3 70B) Pandas Streamlit
Real finding
An 8B model couldn't reliably follow the agent's exact output format and looped on errors
 
Switching to a 70B model fixed it immediately — model size affects format-following, not just answer quality

LoRA Fine-Tuning for a Fixed Answer Format

GenAI Completed

Fine-tuned distilgpt2 with LoRA to consistently follow a fixed answer format, training ~0.36% of the model's parameters. Evaluated with a genuine before/after comparison, including questions never seen during training.

Python Hugging Face Transformers PEFT / LoRA Google Colab
Training setup
Trainable params0.36% of model
Training loss~5.2 → <0.1
Key finding
Learned the target format reliably, even on unseen questions
Did not reliably learn new facts — an honest, expected limitation at this scale

Extractive vs. Generative Question Answering

GenAI Completed

A controlled comparison of extractive QA (a BERT-family model that can only point at existing text spans) against generative QA (LLM, grounded via prompting) on the same source text and questions.

Python Hugging Face Transformers DistilBERT / SQuAD Groq API (Llama 3.3 70B)
Headline finding
Extractive QA's confidence score did not reliably predict correctness
 
Wrong answer at moderate confidence (0.43); empty answer at high confidence (0.92) — the worse failure mode of the two
📦 Additional Projects
RAG Chatbot for User-Uploaded PDFs GenAI

The same retrieval pipeline as the papers project, generalized to arbitrary, user-uploaded documents rather than a fixed, curated knowledge base.

Stable Diffusion Prompt Engineering Comparison GenAI

A controlled test of negative prompting and style-modifier keywords on Stable Diffusion, holding the subject constant. Negative prompting produced a clear, visible improvement; style keywords showed little effect on this endpoint — a real, somewhat counterintuitive finding.

SMS Spam Classifier (Classical ML)

A deliberate contrast to the GenAI projects above: TF-IDF + Naive Bayes, included to demonstrate that not every NLP problem needs an LLM. Evaluated with precision/recall rather than accuracy alone, given an 87/13 class imbalance.

📊 Data Analytics & Machine Learning

Healthcare Insurance Cost Analysis & Prediction

Completed

A complete end-to-end Data Analytics + Machine Learning project analysing 1,300+ healthcare insurance records to identify key cost drivers and predict medical charges, from raw CSV to deployed Streamlit web app.

Python Pandas Scikit-learn Power BI SQL Streamlit
Key insight
Smokers incur significantly higher charges — the single strongest cost predictor
 
5+ regression models compared; best deployed via Streamlit for real-time estimation

Work history

Full Stack Data Science & GenAI Trainee

Naresh i Technologies

Dec 2025 – Present

  • Built and evaluated RAG pipelines, multi-agent systems, and LoRA fine-tuning projects across LangChain, CrewAI, and Hugging Face
  • Built and evaluated supervised ML models for classification and regression across multiple domains
  • Developed GenAI applications using LLMs (OpenAI, Gemini, Groq), LangChain/LangGraph, and RAG pipelines
  • Applied NLP techniques for text processing, embedding generation, and semantic search
PythonLangChainCrewAIHugging FaceScikit-learnGit

Software Engineer

Cortracker Inc

Feb 2025 – Nov 2025

  • Designed and optimized complex SQL queries and stored procedures for reporting
  • Built internal dashboards for business stakeholders to track KPIs
  • Developed and maintained REST APIs to enable data flow across systems
  • Collaborated cross-functionally to translate requirements into analytics solutions
SQL ServerASP.NET CoreC#AngularAzure DevOps

Get in touch