Probabilistic forecasting for production signals. inertialai-forecast returns low-latency calibrated quantiles from signal history. inertialai-forecast (reasoning) adds a reasoning layer when calendar events, inspection images, logs, or analyst notes should change the forecast.
inertialai-forecastTime-Series Only
Calibrated quantiles from the time-series alone. Built for dashboards, automation, alerting, and high-volume jobs where history is enough.
inertialai-forecast (reasoning)Context-aware
Reads text, structured context, and images, then applies a traceable adjustment when outside events should move the forecast.
Promotion starts Monday, supported by a 35% increase in ad spend.
Near-term demand should lift above the signal-only forecast.
On the official GIFT-Eval suite (97 tasks, error normalized by the seasonal-naive baseline), inertialai-forecast posts the best MASE among the listed benchmark systems.
The reasoning layer reads your context and any attached images, then applies a single audited adjustment to the forecast. On controlled scenarios it selects the correct operation 100% of the time — versus 57% for hand-written rules — and delivers roughly 4× lower CRPS.
It also improves a plain forecast with no text at all: by computing properties of the series — non-negativity, integer counts, material recent volatility — it repairs calibration the base model leaves behind, lowering CRPS 1.49% on GIFT-Eval and tightening interval coverage (0.737 → 0.774). It can also explain a forecast in plain language. Send text or an image and the request routes here automatically.
MASE and CRPS on GIFT-Eval, normalized by the seasonal-naive baseline, against a set of public foundation-model baselines.
Controlled scenarios with a known required operation: promotions scale up, outages shift down, stockout constraints clamp negatives, and uncertainty events widen intervals.
PyTorch eager, Triton HTTP, and Triton gRPC paths across batch sizes and client concurrency. gRPC batching is the efficient backend for throughput-oriented calls.
The forecaster is InertialAI's own model. The reasoning layer runs on a hosted LLM that you choose — pick for cost or capability. The model you select is the model that runs: there is no silent fallback.
Recurring forecast jobs, alert thresholds, capacity planning, and high-volume inference where signal history is enough.
When the forecast depends on external context such as launch calendars, maintenance notes, market commentary, images, or operational constraints.
