I'm a Software Engineer – AI/ML at Oracle focused on building production AI systems across LLM applications, retrieval, model evaluation, inference optimization, and cloud infrastructure.

My background combines:

• Production AI/ML: RAG, LLM evaluation, fine-tuning, model serving, inference optimization
• ML Systems: PyTorch, distributed training, CUDA, NCCL, FSDP, vLLM, FlashAttention
• Distributed Systems: microservices, replication, consensus, fault tolerance, Kubernetes
• Backend Engineering: Python, Go, Java, REST APIs, PostgreSQL, Redis
• Research: LLM-based path planning, digital twins, wireless connectivity optimization

I enjoy working at the intersection of machine learning and systems engineering — taking models from experimentation through optimization, deployment, evaluation, and production operation.

Jun 2025 – Present | United States

• Engineer production ML/LLM systems spanning data processing, retrieval, ranking, evaluation, and inference.
• Built enterprise RAG pipelines covering document processing, embeddings, hybrid search, reranking, and LLM generation, improving retrieval relevance 21% and reducing hallucinations 30%.
• Developed Python/SQL pipelines and automated evaluation frameworks measuring accuracy, retrieval quality, ranking performance, latency, and failure rates across thousands of queries.
• Applied SFT/LoRA fine-tuning, batching, KV caching, and quantization, improving task accuracy 78% → 86% and reducing p95 latency 1.2s → 420ms.
• Productionized ML/GenAI services using AWS, Docker, Kubernetes, and CI/CD with validation, monitoring, versioning, and rollback.

Aug 2024 – May 2025 | Arlington, TX

• Developed an LLM-powered connectivity-aware navigation system combining OpenStreetMap/OSMnx, Blender Digital Twins, and NVIDIA Sionna for realistic mmWave signal propagation.
• Automated Digital Twin dataset generation into 1,000+ instruction-tuning samples using Python, NumPy, and Pandas on 4× NVIDIA A30 GPUs.
• Fine-tuned GPT-4o Mini on routing datasets generated with a modified Dijkstra algorithm, improving valid route generation 71.4% → 100% while achieving 76.4% oracle signal coverage.
• Accelerated routing experiments ~60% using Dockerized GPU environments and PyTorch/CUDA across 10K+ routing scenarios.

Jun 2019 – May 2023 | Chennai, India

• Scaled Java/Spring Boot microservices on OCI and Kubernetes from 20 to 50+ services across development, staging, and production.
• Owned production services and resolved 20+ critical incidents, improving SLA compliance from 95% → 99.9%.
• Increased distributed-system capacity 40% while reducing infrastructure overhead 20% through Spring Boot, PostgreSQL, Redis, and connection-pool optimization.
• Automated deployments using Jenkins, GitHub Actions, Docker, and Kubernetes, enabling blue-green deployments and zero-downtime releases.
• Conducted 100+ code reviews and mentored 5+ engineers while improving API response time 25% and reducing database load 20%.

IEEE Open Journal of the Communications Society — Accepted 2026

Multi-objective path-planning framework combining wireless digital twins with a fine-tuned LLM for connectivity-aware navigation in 6G/XR environments.

• 1.9% outage probability vs. 2.3% for RL baseline
• 312 ms average inference latency
• Tunable optimization parameter controlling signal strength vs. travel distance
• GPT-4o Mini fine-tuned using Digital Twin-grounded routing data
• Deterministic sanitization improved raw LLM path validity 65% → 100%

Stack: GPT-4o Mini, OpenAI Fine-tuning API, NVIDIA Sionna, OpenStreetMap, Blender, Python, PyTorch

📖 arXiv

IEEE ICC 2026

LLM-driven path optimization for next-generation mmWave/6G networks using Digital Twin simulation and connectivity-aware routing.

• Connectivity-aware routing using mmWave signal propagation
• Digital Twin-based network simulation
• Modified Dijkstra routing
• LLM-based path planning
• Practical edge inference considerations

📖 arXiv

Mechanism: vLLM PagedAttention + CUDA graph capture.

NF4 reduced VRAM 61.8% with approximately 16% throughput loss.

• PyTorch, JAX, Hugging Face Transformers
• LLMs, RAG, Fine-Tuning, LoRA/QLoRA
• SFT, DPO, GRPO, RLVR, Reward Models
• Embeddings, Vector Search, Reranking
• Agentic AI, Tool Calling, Model Evaluation

• PyTorch DDP, FSDP, NCCL
• CUDA, FlashAttention-2, vLLM
• TensorRT, ONNX Runtime
• KV Cache, PagedAttention
• Quantization: NF4, INT8
• Distributed Training, Model Serving

• Go, Java, Python, C++, SQL
• Raft, Replication, Partitioning, MVCC
• Fault Tolerance, High Availability
• Microservices, REST, gRPC
• PostgreSQL, Redis, MySQL

• AWS: EC2, S3, Lambda, SageMaker
• Kubernetes, Docker
• Jenkins, GitHub Actions, CI/CD
• Prometheus, Grafana
• Oracle Cloud Infrastructure

• 21% improvement in RAG retrieval relevance
• 30% reduction in RAG hallucinations
• 78% → 86% task accuracy improvement
• 1.2s → 420ms p95 LLM latency reduction
• 3.14× distributed training throughput improvement
• 21% per-GPU memory reduction with FSDP
• 1.24–1.32× vLLM inference throughput improvement
• 61.8% VRAM reduction with NF4 quantization
• 20 → 50+ production microservices at TCS
• 95% → 99.9% production SLA improvement
• 40% backend capacity improvement
• 100+ code reviews
• 5+ engineers mentored
• 1,000+ Digital Twin instruction-tuning samples
• 10K+ routing scenarios
• 100% valid route generation
• 0 data loss across 100 LedgerDB failure-recovery cycles

• AWS Certified Data Engineer – Associate
• Microsoft Certified: Data Engineer Associate
• Oracle Cloud Associate Cloud Engineer
• Oracle AI Vector Search Specialist
• Neo4j Certified Associate

I'm interested in opportunities across AI/ML Engineering, ML Systems, LLM Infrastructure, Distributed Systems, and Software Engineering.

• 💼 LinkedIn
• 💻 GitHub
• 📚 Google Scholar
• 🌐 Portfolio