Devansh Tayal
I am a passionate developer focused on building AI models, full-stack applications, and exploring new technologies. Lately that means ML security from a systems angle: whether a security claim still holds once a model is connected to tools, state, and real workflows.
I love creating interactive experiences that center people and their feedback in everything I build. A core part of my philosophy is the belief that every piece of advice, big or small, is valuable, and I'm always eager to learn and grow from it. At home, I enjoy spending time with my family, especially my lab, who's always by my side. I also love connecting with people from different perspectives, watching movies and shows, listening to music, and building new projects.
Projects
Research & Systems
CUDA Tile-Based Renderer
A CUDA renderer and scan pipeline that uses tile binning, prefix-sum scanning, and ordered blending to handle millions of circles efficiently, with performance analysis on an RTX 2000 Ada Generation GPU.
- CUDA
- C++
- Thrust
- +2 more: Parallel Scan, GPU Rendering
Web Development & AI
potentially
A calendar prototype that stores tasks and completion state in Supabase, with an experimental scheduling layer for detecting recurring habits and suggesting the next event.
Game Development
EscapeMate
A hackathon-built escape room assistant that turns room constraints into a playable story, narrates each stage, answers questions, and advances when the physical puzzle reports success.
Other Miscellaneous
Night Utility Framework
A high-performance Roblox automation framework adopted by 50,000+ users, including 10,000 weekly active users.
- Lua
- Reverse Engineering
- UI Development
- +2 more: C#, C++
Experience
Software Engineering Intern
May 2026 - August 2026Auto-Owners Insurance · Internship
- Built reusable .NET 8 and Python/FastAPI health-check packages standardizing a shared readiness contract with HTTP, SQL, and custom dependency checks, parallel execution, bounded timeouts, caching, and secret-safe error handling; delivered across two application stacks.
- Designed Angular, ASP.NET Core, EF Core, and T-SQL workflows surfacing dependency-level health, mapped ETL execution, and permission-aware warnings; integrated existing job-monitoring data rather than rebuilding it, keeping live, historical, and manual signals distinct.
- Separated liveness and readiness probes so downstream dependency failures surface as diagnostics instead of triggering pod restarts.
- Built a distributable code-quality toolchain for Angular/.NET repositories with a Git-scoped C# checker intersecting diff hunk ranges with Roslyn formatter output, enforcing standards on changed lines while grandfathering legacy formatting instead of producing repository-wide mechanical diffs.
AI & SWE Engineer
June 2024 - March 2025Digital ConsultingHaus LLC / GmbH · Part-time
- Containerized and deployed scalable vision & language models on Google Cloud Run and RunPod, cutting inference latency and cost by ≈30%.
- Fine-tuned OpenCLIP and BLIP for product classification & captioning.
- Compressed models via post-training quantization & pruning to fit edge footprints while retaining quality.
- Implemented a voice-command layer for an MDM chatbot, enabling hands-free inventory queries.
Professorial Assistant
Aug 2023 - April 2025OPTML Lab, Michigan State University · Part-time
- Investigated adversarial robustness of Vision Transformers, probing LoRA and Visual Prompt Tuning backdoors with custom PyTorch pipelines.
- Demonstrated 90%+ attack success on LoRA-trained ViTs across GTSRB, Flowers102, and CIFAR-10.
- Rebuilt trigger mechanisms for VPT, surfacing shared failure modes across fine-tuning strategies.
- Benchmarked and tested various novel jailbreak strategies on multimodal models (DeepSeek Janus-Pro) to map cross-architecture vulnerabilities.
Skills
AI & Machine Learning
Frameworks & Tools
Education
Columbia University
in the City of New YorkM.S. Computer Science
Michigan State University
Honors CollegeB.S. Computer Science
Relevant coursework