Gramm started with two papers.
In late 2025, SC Hong and I published a benchmarking study on deep learning architectures for grid demand forecasting (arXiv:2602.21415). We evaluated transformers, temporal convolutional networks, and recurrent architectures against the day-ahead forecasts published by seven U.S. ISOs. The results were clear: modern architectures reduced MAPE by 25-71% depending on the region. The best models ingested raw weather covariates — temperature, humidity, wind speed, solar irradiance — rather than relying on hand-engineered features.
The second paper focused on state space models for safety-critical energy systems (arXiv:2601.01410). We showed that structured state space layers could match transformer accuracy while providing more stable inference under distribution shift — exactly the kind of robustness you need when a forecast is feeding into grid operations decisions during a winter storm or heat wave.
After publishing, the question we kept getting from industry contacts was straightforward: where can we get this? Not the paper. Not the model weights. A production API — something that returns hourly forecasts with sub-second latency, handles versioning, and ships with an SLA.
The answer was: nowhere. There was a gap between published research results and available commercial products. The ISOs publish their own forecasts, but those are regression-based or simple neural nets calibrated for normal conditions. The few commercial forecast vendors in the market use proprietary methods with no published benchmarks. No one was taking the architectures proven in the academic literature and offering them as infrastructure.
That gap became Gramm. We incorporated in Delaware in early 2026 and started building the API. The first question was scope: which grids, which horizons, which delivery format. We chose to start with all seven major U.S. ISOs — CAISO, ERCOT, PJM, MISO, NYISO, ISO-NE, and SPP — because the benchmarking paper already had trained models and evaluation pipelines for each. Day-ahead through 15-day horizons. REST API with JSON, CSV, and XML output.
The hardest part was not the modeling. It was the infrastructure: ingesting weather data from multiple providers in real time, running inference on a schedule that hits delivery windows, handling model updates without downtime, and monitoring accuracy continuously against actuals. Research code does not do any of this.
We launched with 7 grids, published the benchmark results on the website, and made the accuracy page show live MAPE tracking. The philosophy is simple: if you cannot inspect the numbers before buying, the product is not ready.
That is where we are. Two papers became an API. The research is public, the benchmarks are public, and the product is live.