Devpost
Participate in our public hackathons
Devpost for Teams
Access your company's private hackathons
Grow your developer ecosystem and promote your platform
Drive innovation, collaboration, and retention within your organization
By use case
Blog
Insights into hackathon planning and participation
Customer stories
Inspiration from peers and other industry leaders
Planning guides
Best practices for planning online and in-person hackathons
Webinars & events
Upcoming events and on-demand recordings
Help desk
Common questions and support documentation
Sort by:
A Catalogue-Grounded Shopping Copilot Powered by Intent Transparency
Shape-aware Transformer optimization delivers a measured 4.98× geometric-mean speedup across 13 test shapes and validates full 100,000-token sequences on an H100 GPU using independent accuracy checks.
Perry is an accuracy-gated GPU Transformer that delivers 11.8x speedup across 13 official workloads and makes a 100K-token case run within 32 GiB on an RTX 5090.
Compression and reposting change an image's pixels but not its origin. We fine-tune a vision encoder so its features ignore that damage, and detection accuracy holds where it usually drops.
11.3× faster transformer inference on a laptop GPU.
The autonomous agent framework for GPU kernel optimization: The agent does the exploring, tuning, measuring and record-keeping all in a end-to-end loop.
Automated GPU kernel research — compile, benchmark, optimize, repeat, honestly.
We split the 14 given test cases into three families. The first two families dispatch to one bespoke megakernel each. The last family is handled by one of three highly optimized hybrid strategies.
An LLM agent that diagnoses, changes, retrains and measures a recommendation model in a loop: 0.6016 → 0.6056 on KuaiRand-Pure, no human intervention, and a harness built to check its work.
Achieved 1.9x to 31.7x Performance Increase, with 7.7x Geomean on a single-node RTX 4090 with Ada Lovelace Architecture!
Shape-aware Triton kernels that fuse Transformer attention into tiled, memory-lean GPU launches. 1.977× faster inference on RTX 5070 Ti, zero failed elements across 938M+ outputs.
Optimising the graphical “blocks” of floorplans to make spaces more accessible, connected, and inclusive.
Distraction is a custom implementation of the transformer layer optimised for local hardware. Running up to 46× faster than the baseline.
DUET — Depth-Uniform Ensemble with early Terminate
Ultra-fast AIGC detection on the edge, accessible to all
ByteRush unleashes autonomous AI agents to build smarter recommenders—turning raw behavior data into tested ideas, faster experiments, and personalized experiences with minimal human effort.
A Transformer bottlenecks differently at every input shape. We time the candidates at runtime and cache the winner per shape: ~10x geomean speedup, 13/13 correct, on a laptop GPU.
TraceLens is ByteCop’s robust AI-generated image detector, built to remain reliable after real-world transformations and scale efficiently for platform use.
TICK TOCK, time is compute ⏱️ Baseline? Just the starting line 🏁
A dynamically dispatched, hardware-aware Transformer kernel suite that achieves 11.97x speedups peak and processes 100k tokens by exploiting NVIDIA architecture and custom Triton-Gluon fusion.
A robust AI-image detector designed to withstand real-world transformations and cross-domain shifts, with reproducible evaluation and an interactive demo.
a kernel warper for OPENCASCADE that solves the uuid topology problem for CAD and helps agents work with and interact with these easily with inbuild PINN suit that run FEA and CFD sims in millisec.
Deshi Digest is an AI-powered nutrition platform for Bangladeshi food, now enhanced with WebMCP so AI agents can search, compare, analyze, and help plan healthier local meals.
Perceive. Understand. Predict. Protect.
