🤓 Latest Submissions
QuantGate
QuantGate helps ML engineers decide whether a smaller activation format is ready to release. A single prompt defines the dataset, encoded activation budget and maximum permitted accuracy loss. Vultr Serverless Inference plans the evaluation and generates a constrained Python scoring kernel. A disposable Docker worker on Vultr Cloud Compute compiles the kernel to native C++, verifies calibration results against an independent Python reference and returns measurable evidence. Before code generation, QuantGate checks exact sample overlap across data partitions. It freezes the candidate on development data before evaluating holdout slices. The synthetic sensor demo compares FP32, FP16, BF16 and INT8 with FP32 as the reference. INT8 uses 48 encoded activation bytes per sample versus FP32's 192, but its low-margin holdout accuracy falls to 59.38%, so the gate blocks release. The report shows source code, fingerprints, per-slice measurements and the execution trace. The sandbox disables networking, runs as a non-root user with a read-only root filesystem, and limits CPU, RAM and runtime. A timeout probe demonstrates worker termination and verified container removal. The agent explains the measured verdict after execution. Format rounding is simulated at layer boundaries, and byte counts cover encoded activation values rather than total process memory. The public demo also includes an Optical Digits classifier and an injected-overlap scenario.
8 Nov 2026
