Jiten Bhalavat
AI / ML Engineer
I always start by understanding what the user actually needs, study the problem deeply, and then build solutions that truly solve it end-to-end. I work on AI systems like RAG pipelines and LLM fine-tuning, with a focus on building agents that can reason and take actions.
I’m also exploring post-training and scalable inference, while building voice AI and fine-tuning open-source models to compete with larger systems.
Currently: MS in Applied Machine Learning at University of Maryland, College Park
Previously: ML Engineer Intern at Plutomen Technologies · Research Assistant at CHARUSAT
My Best Past Work
Built RAG pipelines for document intelligence with ~89% data extraction accuracy. Used LLM-as-a-judge evaluation to cut hallucinations and improve relevance, and shipped low-latency (<200ms) inference APIs.
accuracy
latency
Built transformer models using PEFT (LoRA), reducing parameters by 99% and speeding up training by 60% on multi-GPU setups, while improving multilingual code intelligence with strong evaluation (CodeBLEU) and scalable pipelines.
supported
approach
Projects
InterviewAI — Speech-to-Speech Mock Interview Platform
FeaturedAI-powered mock interview platform combining real-time speech-to-speech AI with expert-led sessions. Uses ElevenLabs for voice synthesis, GCP for backend, and Firebase for auth — giving job seekers AI-driven practice and expert feedback in one place.
- Voice pipeline: Real-time Speech-to-Speech using ElevenLabs for natural conversational AI interactions
- Cloud backend: Google Cloud Platform for scalable hosting and ML inference serving
- Auth & DB: Firebase for authentication and real-time data synchronization
- Frontend: TypeScript with responsive UI for a seamless interview experience
- AI evaluation: LLM-based feedback on answer quality, clarity, and communication structure
Finetuning Llama 3.1 8B with GRPO
NewFine-tuned Llama 3.1 8B for math reasoning using GRPO reinforcement learning + LoRA, achieving 78.5% GSM8K accuracy. Enabled efficient consumer-GPU training via 4-bit quantization without quality loss.
- Training method: GRPO (Group Relative Policy Optimization) for RL fine-tuning with verifiable math rewards
- Parameter efficiency: LoRA adapters for lightweight fine-tuning on a 4-bit quantized base model
- Reward design: Custom reward function for math problem verification using symbolic equivalence checking
- Evaluation: GSM8K benchmark achieving 78.5% accuracy — competitive with much larger models
- Hardware: Consumer GPU training via 4-bit QLoRA quantization — no expensive A100 needed
Latest Blog Posts
Understanding LLM Quantization: Why FP32, FP16, BF16 and INT8 Matter for Modern AI Systems
How floating-point and INT8 formats shape model quality, speed, and memory when you deploy and scale modern LLM systems.
Think You Know RAG? Take This 16-Point Checklist and Prove It Wrong
A comprehensive 16-point checklist to evaluate and improve your RAG pipelines. Most people get at least 5 of these wrong.
Proven Techniques to Accurately Parse Your PDFs
PDF parsing is a bottleneck for most RAG systems. These techniques improve extraction accuracy and downstream retrieval quality.