Three missions, indexed by the regime they live in.

Helios — Solar Flare Lead-Time Forecasting

The Sun produces, on average, a hundred or so M-class flares per active year and a handful of X-class events. The operational question is rarely whether a flare will happen this week — that probability is close to one whenever there is a sizeable active region on disk — but whether one will happen in the next 30 to 180 minutes, where space-weather operators still have time to act. Helios is the Solstice mission that tries to provide that answer.

The training data is drawn primarily from SDO/HMI line-of-sight magnetograms and the AIA EUV imager, paired with GOES X-ray flux histories used to label flare events. The architecture is a temporal U-Net over magnetogram patch sequences, with a separate convolutional branch over AIA 131 and 193 Å images, fused at the bottleneck. The loss is a weighted binary cross-entropy with a calibration head trained on top via temperature scaling, so that the model output is a probability in the operational sense rather than a logit.

Current benchmark: 0.81 True Skill Statistic at the 60-minute M-class horizon on a held-out 2017–2018 evaluation window. Enter the Helios mission page →

Photon — Vision in the Photon-Limited Regime

When a telescope is asked to image a faint target — a high-redshift host galaxy, a brown dwarf in a wide-field survey, the optical counterpart of a gravitational-wave event — the pixels coming off the sensor are dominated by Poisson statistics rather than Gaussian noise. Standard denoising assumptions (BM3D, Noise2Noise on natural images, simple Gaussian-prior diffusion) leak structure into the dark patches and erase faint signal in the bright ones. Photon is the mission that rewrites those assumptions.

Photon ships three components: a self-supervised denoiser trained directly on the instrument’s own data with a blind-spot architecture and a Poisson-aware reconstruction loss; a cosmic-ray segmenter that replaces the canonical LACosmic baseline with a small U-Net trained on hand-labelled streaks; and a faint-source detector that operates on the denoised + cosmic-cleaned image and produces a catalogue with per-source significance.

Benchmark: 1.7× effective gain over BM3D on EMCCD twilight series at EM gain g = 200; F1 = 0.962 on cosmic-ray segmentation. Enter the Photon mission page →

Sky — Survey Transient Pipeline

Wide-field surveys generate alert streams measured in millions of candidates per night. The overwhelming majority of those candidates are not real astrophysical transients but artefacts: subtraction residuals around bright stars, cosmic rays missed by the cleaning step, satellite trails, ghosts from internal reflections, hot pixels. Sky is the mission that distinguishes the few real transients from the many bogus ones, in real time, before any human looks at an image.

The Sky pipeline is built around an image-differencing front end (PSF-matched template subtraction), a real-bogus convolutional classifier trained on millions of labelled stamps, a host-galaxy association step against external catalogues, and a light-curve fitting back end that produces a calibrated transient type prediction for downstream brokers. The whole pipeline is benchmarked on replayed ZTF alerts and is being adapted for the Rubin alert format ahead of LSST operations.

Benchmark: 97.4% real-source recall at 0.5% bogus contamination on the 2024 January ZTF replay. Enter the Sky mission page →

Why these three and not others

Every choice to take on a mission is a choice not to take on three other plausible ones. We are aware, for example, that exoplanet transit detection, gravitational-wave electromagnetic-counterpart imaging, and radio-interferometric imaging all have substantial computer-vision content and active client communities. We do not work on them today because the company is small and the three current missions already consume our engineering bandwidth. We expect to expand the roster as the team grows, and we publish our shortlist of candidate next-missions on the Dispatches page each quarter.

The unifying property of the current roster is that each mission has (a) a publicly available training corpus large enough to train a meaningful model without ad-hoc data collection, (b) an instrument community that already understands what a calibrated probability means, and (c) a downstream user who can act on the output in operational time. We are not interested in missions that lack any of these three.