Information Technology graduate, first-author vision-language model evaluation researcher, and open-source patch author building scalable, resource-efficient AI systems. My work connects measurable research, distributed GPU infrastructure, and production-grade software engineering.

GitHub · LinkedIn · ORCID

Under review, 2026 for the NeurIPS 2026 Workshop on Grounded and Faithful Vision-Language Models, VLM4RWD, Sydney.

• Traced the tokenizer prefix-space artifact in POPE yes/no extraction and corrected the logit read-out, lifting the true LLaVA-1.5 baseline by 6.13 F1 points to 82.21 F1.
• Re-evaluated six published mitigation methods against the corrected baseline. Several reported gains shrink or reverse after the correction.
• Confirmed the confound across 4 VLMs, 3 tokenizer families, 9,000 queries, and 90 GPU-hours. Attention probing measured Cohen's d = -0.46.
• Released pope-audit on PyPI, version 0.1.1, plus 9,000 prediction records on Hugging Face for reproducibility. The source is archived on Zenodo with a DOI.

Published, March 2026 at ICCET 2026, Chennai. ISBN: 978-81-986418-1-6. Proceedings are to be indexed in IEEE Xplore and SCOPUS.

• Designed a hybrid cloud-edge Platform-as-a-Service where a FastAPI priority scheduler routes CUDA workloads through a Tailscale VPN mesh to cloud GPUs and Raspberry Pi 4B edge nodes.
• Built a Next.js dashboard that streams live worker telemetry through Server-Sent Events.
• Measured queue wait, execution, and network components across tensor payload sizes, with pytest integration tests, Playwright end-to-end tests, JWT authentication, and Dockerized deployment.

FastAPI React GPT-5.6 Terra MCP

• Built a repository preflight engine that orchestrates four parallel GPT-5.6 processes limited to read-only workspace tools, mapping module boundaries, risk zones, and test coverage before autonomous code execution.
• Implemented an application-layer firewall that silently drops model outputs failing strict file-path and line-range validation. Exposed the verified pipeline through a React dashboard, CLI, and local stdio MCP server.

Kotlin BLE mesh LoRa FastAPI On-device LLM

• Built an off-grid SOS network where Android phones relay compressed voice reports phone-to-phone over Bluetooth Low Energy, then bridge them over kilometres through a 433 MHz LoRa gateway on Raspberry Pi and Arduino.
• Fine-tuned and released Sahayak-E2B, a Gemma-based model, with QLoRA. Exported GGUF for fully offline triage on a Snapdragon NPU and open-sourced both Base and GGUF variants on Hugging Face.

RAG IBM Granite 3.1 BM25 + embeddings React

• Developed a hallucination-resistant explainer for football VAR calls. A hybrid BM25-plus-vector retriever over 593 rule chunks feeds IBM Granite 3.1 8B under a strict JSON schema and abstains when evidence is weak.
• Validated it on a 50-question adversarial suite. It abstains on out-of-scope queries instead of fabricating and scored full marks on citation, keyword, and decision-type checks.

PyTorch TensorRT

• Extended a faculty-designed YOLOv10-Nano, 96.2% mAP@0.5, and CBAM-MobileNetV2, 93.6% accuracy, pipeline to 5,010 images.
• Executed INT8 TensorRT layer fusion for sub-100 ms Jetson Nano latency under 50 MB.

• National Finalist, Snapdragon Multiverse Hackathon, Bengaluru, July 2026, for Sankat-Mochan.
• Runner-Up, 2nd place, IBM SkillsBuild "AI Inside the Match" International Challenge, June 2026. Awarded $1,250 among 3,381 registrants and 307 projects, judged by a 19-member IBM panel.
• Research Collaborator, MSME Government Grant, 2025. Faculty-invited to design machine-learning algorithms for a sanctioned social-media fake-ID-detection project worth INR 13.5L, about US$14,000.

MongoDB Docker Jenkins GitHub Actions

• Automated GitHub Actions deployments, cutting release cycles by 25% and removing manual steps.
• Built live MongoDB analytics dashboards for demos and ran Docker and CI/CD reliability checks that held 97.8% test-phase uptime.

Java Spring Boot AWS EC2

• Integrated analytics modules into Java REST APIs and tuned Spring Boot with EC2 deployment strategies.
• Improved backend response time by 18% and cut server error rates by 12% under sustained load testing.

B.Tech, Information Technology | Sri Krishna College of Technology, India | October 2022 to April 2026 | CGPA: 8.41 / 10

• Sri Krishna College of Technology is autonomous, NAAC A, and NBA-accredited.
• Awarded 10/10, O, Outstanding in the Semester 8 capstone, Project Work Phase II, for VisualPC, with a cumulative top-5% department cohort rank.
• Relevant coursework: Artificial Intelligence, Data Visualization, Computer Networks, Cloud Computing and Containerization, Data Structures and Algorithms, and Operating Systems.
• NVIDIA DLI, Deep Learning and AI on Jetson Nano, 2024.
• IEEE Student Member and IEEE Computer Society Member.

GitHub · LinkedIn · ORCID