What should happen when the LLM in your product is wrong, or the API is down? I built GuardRail to answer that for infrastructure code. Seven deterministic rules parse a Terraform plan and decide pass or fail. Claude only explains the risk and drafts the corrected HCL, so the verdict still ships when the API fails.
I also know what breaks inside the models. I attacked LLaMA 3 adapted for image classification with FGSM across 60K MNIST examples and watched accuracy fall from 83% to 66% once the adaptation disrupted self-attention. In a separate project I merged two separately fine-tuned BERT models into one 110M-parameter model by composing task vectors. It reached 87% on sentiment and 74% F1 on NER with 50% less storage.
The engineering underneath is production work. At Persistent Systems I interviewed 8 developers about why code-review findings were ignored, then built the dashboard around their answers. It covered 15+ enterprise codebases and cut vulnerabilities 30%. I also shipped 10+ Spring Boot APIs and moved deploys from hours to minutes with Docker-based CI/CD. At Virginia Tech, I was the primary developer on a live exam platform where I built the exam-management module from scratch, made the frontend WCAG-compliant for 2,000+ students with registered disabilities, and users grew 42% to 7.2K+.
Looking for software engineering roles building LLM products, where I can own a service from design to production. MS in Computer Engineering from Virginia Tech, available immediately.
Remote: Yes, (open to onsite/hybrid)
Willing to relocate: Yes
Work authorization: Authorized to work in the US on OPT (STEM OPT extension eligible)
Technologies: Python, PyTorch, Hugging Face, Claude/OpenAI APIs, LLM tooling, Java/Spring Boot, TypeScript/Node.js, React, MySQL, Docker, Kubernetes, Terraform, GCP (Associate Cloud Engineer)
Résumé/CV: https://drive.google.com/file/d/1KdL-4_1AtpJYFj6JQIG3xkNYCz4...
GitHub: https://github.com/ishani02
Portfolio: https://ishanikohli.vercel.app
Email: [email protected]
LinkedIn: https://linkedin.com/in/ishani-kohli
What should happen when the LLM in your product is wrong, or the API is down? I built GuardRail to answer that for infrastructure code. Seven deterministic rules parse a Terraform plan and decide pass or fail. Claude only explains the risk and drafts the corrected HCL, so the verdict still ships when the API fails.
I also know what breaks inside the models. I attacked LLaMA 3 adapted for image classification with FGSM across 60K MNIST examples and watched accuracy fall from 83% to 66% once the adaptation disrupted self-attention. In a separate project I merged two separately fine-tuned BERT models into one 110M-parameter model by composing task vectors. It reached 87% on sentiment and 74% F1 on NER with 50% less storage.
The engineering underneath is production work. At Persistent Systems I interviewed 8 developers about why code-review findings were ignored, then built the dashboard around their answers. It covered 15+ enterprise codebases and cut vulnerabilities 30%. I also shipped 10+ Spring Boot APIs and moved deploys from hours to minutes with Docker-based CI/CD. At Virginia Tech, I was the primary developer on a live exam platform where I built the exam-management module from scratch, made the frontend WCAG-compliant for 2,000+ students with registered disabilities, and users grew 42% to 7.2K+.
Looking for software engineering roles building LLM products, where I can own a service from design to production. MS in Computer Engineering from Virginia Tech, available immediately.