Skip to main content

​
GPU Kernel Optimization

Write custom CUDA/Triton kernels for 10 KernelBench problems spanning element-wise ops, fused operator chains, and full transformer blocks. Each solution must produce correct outputs and run faster than the PyTorch reference. What made the difference: The KB agent discovered TF32 tensor cores, a hardware feature that doubles matmul throughput with no precision loss for inference. This gave +110% speedup on transformer blocks (3.67x vs 1.75x). The KB agent also fused Swish+Bias+GroupNorm into a single Triton kernel and wrote custom depthwise conv kernels with float4 vectorization + CUDA Graphs, while the baseline fell back to cuDNN. → Full results and replication instructions

​
LLM Post-Training

Implement a complete SFT + DPO fine-tuning, LoRA merge, vLLM serving, and IFEval evaluation pipeline for Qwen/Qwen2.5-1.5B on 8×A100 GPUs. What made the difference: The KB agent queried hyperparameter priors and chose a 25x lower DPO learning rate (2e-6 vs 5e-5), avoiding divergence during preference optimization. It also merged the SFT adapter into base weights before creating a fresh LoRA for DPO (avoiding stacked adapters), used a proper train/eval split for DPO, and shuffled datasets with a fixed seed. → Full results and replication instructions

​
Self-Evolving RAG

Build a FastAPI service that ingests a corpus, answers questions via hybrid retrieval-augmented generation, and automatically improves itself over multiple evolution rounds by diagnosing retrieval failures, re-chunking documents, and adapting query strategies. What made the difference: The KB agent used DBSF (3-sigma) score normalization for fusing BM25 and vector scores instead of Reciprocal Rank Fusion. This preserved score magnitude information, letting the evolution loop meaningfully adjust fusion weights. It also used a blue-green collection pattern for ChromaDB index swaps instead of a simpler lock-based pointer swap. → Full results and replication instructions

​
Customer Support Agent

Build a multi-agent system that classifies 200 customer support tickets into 27 fine-grained intent categories using agent handoffs, state persistence, and structured output. What made the difference: A single KB lookup taught the agent typed handoff callbacks with input_type, a framework-specific pattern in the OpenAI Agents SDK. The baseline used fragile regex extraction from free text, while the KB agent registered each handoff with a Pydantic model that the SDK enforces at handoff time. The KB also taught proper state serialization and output_type usage. → Full results and replication instructions