Use cases
Where seconds of CPU
beat hours of GPU.
Each use case below is grounded in a measured result — not a vision slide. Where the grounding is a property of the construction rather than a benchmark, we say that instead.
Document & ticket routing
Route support tickets, emails, claims, and complaints to the right team. Synthesized from your labeled history in seconds; re-synthesized every night, or every hour, for the cost of a CPU coffee break.
Grounded in the 20 Newsgroups result: 4-way text routing at 87% — above a tuned SVM.
On-device & edge classification
Closed-form synthesis needs no training infrastructure, so models can be built on the device that uses them — field sensors, kiosks, clinics with no connectivity, no cloud round-trips.
Synthesis runs in seconds on commodity CPUs; the Lab on this site synthesizes models in your browser tab.
Privacy-preserving ML
When data cannot leave the premises — health records, legal files, student records — the model can be grown where the data lives. Nothing is shipped to a GPU cluster.
No training pipeline means no data movement; the entire model is a deterministic function of local statistics.
High-cadence retraining
Spam patterns, product catalogs, and news topics drift weekly. Where teams retrain monthly because training is expensive, synthesis rebuilds the whole model from scratch in under a minute.
Full 20 Newsgroups ensemble synthesis: ≈60 seconds on CPU, end to end.
Auditable, deterministic models
Every weight is a closed-form function of the data — the same corpus always produces the same model. Features are Fisher-scored and inspectable. For regulators, that is a provenance story gradient descent cannot tell.
Determinism and feature inspectability are properties of the construction itself (paper §3).
Cold-start & fallback language models
A ~2.0 bits-per-byte byte-level model, synthesized in seconds, serves as an offline fallback, a sanity-check floor for trained systems, or a drift detector — not a chatbot, and we say so.
Paper §6: the zero-training n-gram floor (2.12 BPB) beats a small GRU trained for 3,000 steps (2.74).
And where we'd turn you away
If you need a generative frontier model, you need training.
Open-ended text generation, dialogue, code synthesis — our own research shows closed-form methods cannot reach that frontier, and we will tell you so before you spend a rupee. What we will build you instead is the classifier, router, or triage model that runs beside it — for a thousandth of the energy.