I build the parts of AI systems that have to be correct.

The LLM returns data, never actions. It hands back a plan; the runtime decides whether that plan is legal — does that tool exist, are the argument types right, is there a cycle in the dependencies — and only then executes it as a parallel DAG. Every output is verified. Anything wrong is repaired under a bounded budget, keeping the steps that worked.

Then I measured what each stage is actually worth by turning it off. The result worth knowing: drop the verifier and the run still reports success: true while returning a wrong answer — only an external grader notices. Drop the repair budget instead and you get the same score but a loud failure. And keeping only the free structural checks matched full accuracy at a third of the LLM calls, which says the LLM judge earned nothing on that suite.

The whole engine — planner, validator, DAG executor, verifier, replanner and the eval harness — is 40 kB of dependency-free JavaScript that runs in your browser. Open the live link, build your own configuration, swap any stage prompt, add your own tasks, and watch the plan graph execute node by node.

• TypeScript
• Agent orchestration
• Tool calling
• DAG scheduling
• LLM-as-judge
• Evals
• Zero deps

FlashAttention-2 forward and backward kernels written in Triton from the papers (arXiv:2205.14135, arXiv:2307.08691). Online softmax with running max and sum statistics keeps tiled attention numerically exact without ever materialising the N×N score matrix in HBM.

Activation memory drops from O(N²) to O(N) by recomputing probability tiles in the backward pass from a saved log-sum-exp vector. Exactness is validated against an fp32 oracle across fp16/bf16, causal masking, head dimensions and ragged sequence lengths — because a fast attention kernel that is quietly wrong is worse than a slow one.

Ships with an autotuned benchmark harness reporting latency, achieved TFLOP/s and peak memory against scaled_dot_product_attention and an O(N²) baseline, plus a drop-in nn.Module.

• Triton
• CUDA
• PyTorch
• GPU kernels
• Attention internals
• Numerical exactness