A training run keeps computing gradients for layers that stopped learning hours ago. EnAi watches gradient energy per layer while the run is in flight, freezes each layer the moment its contribution collapses, and raises the batch size into the headroom that frees up. All of it happens mid-run, with no restart and no change to the model code.
Solo. I build the engine, the study and the measurement.
6.91%
training compute removedAnalytic ledger, confirmed by autograd
+20.9%
throughput once engagedMachine normalised, median of 12 runs
1,563 to 0
backward passes into frozen layersCounted per layer by hooks left attached
92%
of the theoretical ceiling reached+7.42% predicted, +6.85% measured
A small group building AI products around one question: what is still harder than it needs to be? Everything ships free and open source. The first product out is Admission OPS, because university admissions is one of the most paperwork heavy moments in a person's life and most of it is deadline tracking that software should have absorbed years ago.
Peer to peer video with emotion inference in the browser.
A peer to peer video chat on WebRTC, handling signalling and connection state directly, with facial emotion inference running in the browser rather than on a server. The interesting constraint was frame sampling: infer too often and the call stutters, too rarely and the reading is stale.
Product search over data that never agrees with itself.
A product search and comparison tool over external APIs that return the same field in a different shape every time. The work is in the normalisation layer and the caching: filtering and sorting happen against a cleaned local shape, so the interface stays responsive instead of waiting on a round trip per keystroke.
An async audio pipeline that stays responsive on long recordings.
Speech to text, emotion classification and LLM based summarisation over a single asynchronous Python pipeline. Batching and background processing keep response latency flat as recordings get longer, instead of making the user wait for the whole file.
