Looking for a job

Anay Chauhan

@anaychauhan

Applied ML researcher and R&D engineer focused on efficient LLM inference, RAG systems, and production-ready model evaluation.

India

Message

What I'm looking for

I’m looking for a team where I can build and ship efficient LLM inference and agentic RAG systems—blending research rigor with production evaluation, correctness checks, and deployment-focused engineering on real-world models.

I’m an Applied Machine Learning Researcher and R&D Engineer building efficient, memory-aware capabilities for transformer-based LLM inference. I designed a kernel-native KV cache compression method that couples Angle Domain Attention with a rate-distortion objective, achieving 1.55–1.72× higher decode throughput at matched perplexity and reducing HBM bytes/token by 24% through a compressed-domain GPU kernel without dense-key reconstruction.

At Synopsys, I shipped an agentic spec-to-code Retrieval-Augmented Generation (RAG) pipeline using a DSL object model, and I fine-tuned Qwen2.5 with LoRA adapters to improve DSL schema compliance by 35% over prompt-only baselines. I also architected AST-level correctness checks, schema validation, and regression-test infrastructure to reduce silent failures, and I delivered a FastMCP server integrating Coverity static analysis for automated code-review filing adopted across 100+ engineers.

Experience

Work history, roles, and key accomplishments

Current

Applied Machine Learning Researcher

Current

Pragya Labs

Jan 2026 - Present (6 months)

Designed a kernel-native KV cache compression method using Angle Domain Attention with a rate-distortion retention objective, achieving lossless-equivalent quality at reduced cache sizes for LLM inference. Improved decode throughput by 1.55–1.72× at matched perplexity on 8K/32K/128K windows and reduced HBM bytes/token by 24%.

LLM Inference Angle Domain Attention Rate Distortion Optimization GPU Kernel Optimization

Current

R&D Engineer

Current

Synopsys

Jun 2025 - Present (1 year 1 month)

Designed and shipped an agentic retrieval-augmented generation (RAG) pipeline for spec-to-code synthesis using a DSL object model, deployed in production tooling for Google, Qualcomm, and MediaTek. Fine-tuned Qwen2.5 with LoRA (SFT) to improve DSL schema compliance by 35% and shipped an MCP server integrating Coverity static analysis adopted by 100+ engineers.

Retrieval Augmented Generation (RAG)Spec To Code Synthesis DSL Object Model Qwen2.5 AST Level Correctness Checks Model Evaluation Coverity

Undergraduate Researcher

Translational Biology Lab

Jan 2024 - May 2024 (4 months)

Built ML pipelines for protein-ligand binding (IC50) prediction on 11K ChEMBL samples for ALK and BRAF using Mordred/RDKit descriptors. Improved binding affinity prediction by 20% via cross-feature interaction terms and achieved k-fold cross-validated R=0.92 (Linear Regression) and R=0.89 (Random Forest).

ChEMBL Datasets RDKit Mordred Features Cross Feature Interaction Modeling Machine Learning Pipelines Random Forest Cross Validation