Srikar Veluvali

Software engineer and recent KMIT graduate working across AI systems, compilers and high-performance computing.

About

I am a recent Information Technology graduate from Keshav Memorial Institute of Technology. I build AI products, developer tools and performance-focused systems.

Lofi programmer working at a laptop beside a rainy window

My experience includes production software at Amazon, compiler research at Microsoft and GPU-accelerated simulation work at DRDL-DRDO.

Education

B.Tech in Information Technology

Keshav Memorial Institute of Technology, 2022-2026

Academic record

Branch Topper, CGPA 9.53

Ranked first in the IT branch across all semesters

Experience

Jan 2026 - Jun 2026Hyderabad, Telangana

Amazon

Software Development Engineer Intern

Built agentic AI and workflow infrastructure inside Amazon's Case Management System for Controllership teams.

Built automated dispute resolution that detects invoice discrepancies with carriers and drafts auditor-ready resolutions.
Standardized onboarding workflows with engineering and operations to accelerate time-to-productivity for new accounts.
Eliminated a 7-day access SLA by moving Login IDs to LDAP and POSIX groups.
Enabled Controllership teams to self-manage memberships instead of depending on IT for access changes.

Jan 2025 - Jul 2025Bengaluru, Karnataka

Microsoft

Research Intern

Built a graph-based compiler that automatically optimizes NVIDIA PTX assembly without source changes.

Delivered a 74.12% speedup on LLM attention-based matrix-multiplication kernels.
Applied the optimization automatically with no changes to the original CUDA source code.
Designed parsing, SSA conversion, control-flow analysis and symbolic execution stages.
Built a Z3-powered analysis and transformation pass for non-coalesced memory access patterns.

Jun 2024 - Dec 2024Hyderabad, Telangana

DRDL - DRDO

HPC Software Engineering Intern

Ported a production CFD solver from MPI-only C to hybrid MPI + CUDA for missile-flow simulation.

Reduced simulation runtime by 25.93%, from 30 days to 22 days.
Designed and tuned parallel algorithms to increase GPU-core utilization.
Worked with engineers to align compute kernels with fluid-dynamics requirements.
Accelerated delivery timelines through more efficient simulation workflows.

Projects

~/selected-work/grpo-cot-for-qwen-2.5

$ train qwen2.5 --method grpo

GRPO-CoT for Qwen 2.5

A reinforcement-learning pipeline for Qwen2.5-3B using multi-objective rewards, 4-bit QLoRA, vLLM sampling and schema-aligned GRPO training on GSM8K.

output: structured reasoning / single-GPU fine-tuning / 4x improvement

PythonPyTorchQwen2.5GRPOQLoRAvLLM

Source

~/selected-work/agentblocks

$ run agentblocks --self-heal

AgentBlocks

A YAML-driven AI workflow engine with 22 nodes, parallel asyncio execution and an agent-based recovery pipeline that diagnoses and patches failed steps.

output: async DAG execution / retries / repair / versioning

PythonFastAPIReact FlowLLM Orchestration

Source

~/selected-work/prism

$ launch prism --study-mode

PRISM

A personalized learning platform using RAG to turn PDFs, Word documents and YouTube videos into conversations, quizzes, mock tests and study tools.

output: RAG / assessments / interviews / flashcards

ReactFastAPIPineconeMongoDB

Source Live product

~/selected-work/searchly

$ start searchly --agentic-retrieval

Searchly

An agentic product recommendation system that combines web scraping, NLP pipelines, semantic embeddings and Pinecone retrieval in a full-stack search experience.

output: agentic scraping / semantic search / recommendation

PythonFlaskReactPineconeRAG

Source Demo

~/selected-work/astor-ai

$ serve astor-ai --rag