+35% Precision@10
LLMs that check their sources
Multimodal RAG over 100+ page PDFs, text and images, with hybrid search and reranking. Also cut factual hallucinations by 40%.
Open to AI and ML engineering roles
I build LLM systems that stay grounded in the facts.
AI engineer in financial services. I fine tune open models, build RAG over messy enterprise documents, and keep all of it running in production.
Strengths
What I bring to your team
12 → 2 GB model memory
Fine tuning on a budget
QLoRA, PEFT and 4 bit quantization, so domain models train on a single T4 GPU instead of a server room.
AWS + Azure in production
Production, not just notebooks
MLflow, SageMaker and drift monitoring with automated retraining, plus the FastAPI and Java services that serve it all.
Receipts
LLaMA 3.2 3B, before and after my QLoRA fine tune
Trained on 2,658 instruction pairs synthesized from unlabeled company documents. Scored on 766 held out examples.
ROUGE-L
Word overlap with the reference answer
+0.403.9× higher
BERTScore
Meaning match with the reference answer
+0.1928% higher
Skills
The toolkit
Grouped by what they're for. Core marks the ones I use every day.
LLMs and GenAI
Fine tuning, prompting and evaluating language models.
6 core · 11 total
- LLM fine tuning
- QLoRA and PEFT
- Evals
- LangChain
- LangGraph
- Prompt engineering
- Hugging Face Transformers
- LLaMA
- Ollama
- GPT-4
- Claude Code
Retrieval and search
Getting the right facts in front of the model.
2 core · 6 total
- RAG
- Azure AI Search
- Pinecone
- FAISS
- Chroma
- pgvector
Machine learning
Classic and deep learning models.
1 core · 5 total
- PyTorch
- TensorFlow
- scikit-learn
- XGBoost
- LightGBM
MLOps and deploy
Shipping, serving and watching models in production.
2 core · 9 total
- MLflow
- FastAPI
- Weights & Biases
- Kubeflow
- Model monitoring
- Docker
- Kubernetes
- GitHub Actions
- Triton
Cloud
Where it all runs, on AWS and Azure.
1 core · 9 total
- SageMaker
- S3
- Lambda
- EC2
- ECR
- Azure OpenAI
- Azure ML
- Azure Functions
- Blob Storage
Code and data
The languages and stores underneath.
1 core · 11 total
- Python
- SQL
- Java
- JavaScript
- React
- Node.js
- PostgreSQL
- MySQL
- MongoDB
- Redis
- DynamoDB
Experience
The short version
Merrill Lynch
AI Engineer
Jan 2025 to now · Remote
I teach LLMs to read 100 page financial documents and, more importantly, to admit when the answer isn't in there.
- Multimodal RAG on Azure AI Search: +35% Precision@10 and 40% fewer hallucinations.
- Fine tuned LLaMA 3.2 3B with QLoRA and squeezed it from 12 GB to 2 GB so it trains on a T4.
Canadian Western Bank
Software Engineer, AI/ML and Java
Feb 2022 to Jul 2023 · Hyderabad
I kept banking APIs calm during rush hour and taught models to spot fraud, churn and risky loans before they became somebody's very bad Monday.
- Spring Boot microservices that met PCI DSS, SOC 2 and FFIEC requirements.
- Closed 60+ critical UAT defects: 40% fewer bugs reported and 25% better uptime.
Weights & Biases
Machine Learning Engineer Intern
Mar 2020 to Feb 2022 · Hyderabad
I helped data scientists stop losing track of which training run was final_v2_actually_final.
- Built experiment tracking, model versioning and hyperparameter sweeps.
- Shipped SDK integrations for PyTorch, TensorFlow, scikit-learn and Hugging Face.
Projects
Built on my own time
Mar 2025
projhealth.comProj H
Personalized health plans from GPT-4, with an eval suite that stops bad prompts from shipping. It caught a cultural bias bug before users ever did.
250–300 msP95 latency
100–150QPS sustained
20K+datapoints a day
Feb 2025
Case study
AEC conflict detector
Finds conflicting specs across architecture and construction documents, including the drawings, before they become expensive surprises on site.
~70%less manual review
Visionreads the drawings
Fullpytest coverage
Research
Publications and writing
Terahertz Communication for 6G Networks
Coauthored during my MS in Computer Science at the University of North Texas and published by IEEE.
§ Writing
- [2]
Reddy, V. S. “Counting Reel views with HyperLogLog.” Substack. How large platforms estimate unique views with probabilistic data structures.
§ Open source and awards
- [3]
Reddy, V. S. “UI bug report with reproduction steps.” Font Awesome, GitHub issue #19388.
- [4]
Winner, MLH Hackathon. Multilingual AI chatbot built on IBM Watson NLP.
Education
University of North Texas
MS in Computer Science · Aug 2024 to May 2026 · GPA 3.64
Machine learning, artificial intelligence, NLP research, software engineering and computer networks.
Open to new positions
Hiring for AI or ML? Let's talk.
I'm looking for AI and ML engineering roles. Email is the fastest way to reach me.