Clouds are excellent at being clouds. They drift, gather, thicken and ruin the view with complete indifference to a satellite schedule. Night is equally uncooperative. For most of us, this is atmosphere. For anyone trying to understand a flood, a crop, a convoy or an oil field from orbit, it is missing information. Dhenenjay Yadav has arranged his young career around that missing piece.
Yadav is the co-founder and CEO of AxionOrbital Space, a San Francisco company in Y Combinator's Winter 2026 batch. The company works on a peculiar translation problem. Synthetic Aperture Radar, or SAR, can observe the ground through darkness and cloud. Its output, however, resembles the visual world only after expertise and processing have done their work. AxionOrbital is building machine-learning models that convert that radar signal into optical-style imagery intended to be immediately legible to people and ordinary computer-vision systems.
The distinction matters. The space economy has become very good at collecting data. Usefulness still has to survive the trip back to Earth. A sensor may be awake while its customer remains, metaphorically, in the dark.
A space dream with mud on its shoes
Yadav begins his biography with a plain confession: he always dreamed about exploring space and the universe. His path toward that dream ran through agriculture, machine learning and management rather than a launchpad. In 2023, while associated with the College of Agriculture at CCS Haryana Agricultural University, he presented a crop-classification model built with Sentinel-2 multispectral, multi-temporal data. The title is a small forest of technical nouns. The underlying question is wonderfully earthly: what is growing down there?
He later attended the PGP-FABM program at the Indian Institute of Management Ahmedabad from 2023 to 2025. His public biographies identify reinforcement-learning research at IIM Ahmedabad and machine-learning engineering work at the Indian Space Research Organisation. His résumé also contains a KVPY SX Fellowship, UAV building and former competitive PUBG play. The combination is less random than it looks. Each rewards a feel for systems, incomplete information and decisions made before the whole map is visible.
“I always dreamt about exploring space & the universe.”Dhenenjay Yadav
By 2023, Yadav was also a director of Inertialogic AI, an earlier company. His account of that period is important because it supplies AxionOrbital's origin without the usual fog machine. He was dealing with geospatial work that was compute-intensive, repetitive and manual. He checked the frustration with geospatial developers, space startups and space organizations. The problem, he concluded, was shared.
Phoenix learns to molt
When Yadav publicly introduced AxionOrbital in 2025, the product was Phoenix 1.0, a no-code environment for planetary-scale satellite analysis. It promised to let users handle imagery and produce geospatial insight without wrestling every workflow into code. The launch had the useful awkwardness of an early answer: concrete enough to test, unfinished enough to change.
Change came quickly. AxionOrbital later released an open-source geospatial MCP server, connecting language models with satellite, ground-sensor and other data feeds. Yadav framed it with a cheerful question: “Why not Geospatial Agents?” The project pointed toward an agentic environment in which software could assemble and run complex geospatial work.
Then the company pushed toward the visibility problem itself. Hubble, its open model at 10-meter resolution, and Orion, its restricted 0.5-meter model, translate radar imagery into optical-style pictures. The company reports a 0.06-second inference time for Orion and says its deterministic one-step architecture constrains the generated image with spatial information from the input. In this field, a pretty invention is a dangerous one. Operational users need the output to correspond to the ground, not to a model's imagination.
That produces a crisp strategic idea. Instead of putting another sensor in orbit, make the radar sensors already there easier to use. Optical satellites produce intuitive pictures but lose the scene to weather and darkness. Radar keeps working, but demands interpretation. AxionOrbital wants to become the optical layer for that radar economy.
Consider what that layer changes for the person at the other end. A flood-response team does not need a tutorial in radar physics while water is moving through streets. An agricultural analyst should not have to postpone a crop assessment until the next cloudless pass. A commodities desk values a timely count more than an exquisite image delivered after the market has moved. The same translation serves different decisions because the underlying inconvenience is shared: the useful event and the clear optical view rarely consult each other.
AxionOrbital offers enterprise deployment through an API, on premises or at the edge. That last detail is not decoration. Some customers cannot casually send sensitive imagery to a public cloud, and some observations become more valuable when processed close to the sensor. The company's technical story therefore runs from model architecture to the mundane geography of where computation is allowed to happen. Space software eventually meets a server rack.
A laboratory disguised as a company
The public record of AxionOrbital reads like a notebook with the pages still warm. Yadav and researcher Rohan Sawai co-authored DARN in November 2025, a decoder architecture for adapting foundation models to the uneven difficulty of satellite imagery. It adjusts regularization and model capacity by sample, reporting an 86.66 percent mean intersection-over-union result on the multi-task GeoBench benchmark. The paper and company share an instinct: terrain varies, so the machinery interpreting it should adapt.
A conversation recap from AxionOrbital's early period describes an unusually short loop. The team would read a research paper and try to implement it that day. Another account says the founders answered an initial Y Combinator rejection by spending a week improving the model and pressing their case until the decision changed. The charming part is not the mythology of persistence. It is the preference for producing new evidence.
Read the paper. Try it today. Let the result edit the plan.The AxionOrbital operating loop
That loop is expensive when models require serious computing. An interview summary recounts the founders renting a GPU cluster, hunting for credits and continuing while runway was thin. It also records their advice to early founders: ask people for help and do not be shy. Space may be vast; an early-stage network is usefully smaller.
From a picture to a prediction
The next product turn reaches beyond translating what a satellite has already captured. NeoEarth, described publicly in 2026, is a world model for Earth observation designed to forecast how the ground may change up to a month ahead. It maintains an evolving state for a location as new satellite data arrives and conditions its forecasts on climate variables. AxionOrbital says it backtested the model on Hurricane Helene, using the exercise to predict the storm's appearance and trajectory.
The progression is neat enough to sound inevitable only in retrospect. Phoenix asked how more people could analyze satellite data. Orion asked how radar data could become readable. NeoEarth asks what the observed ground might do next. Access, interpretation, prediction. Each product changes the verb.
Yadav's stated long-term aim is broader still: a universal intelligence layer for Earth observation, processing raw satellite pixels so a changing planet becomes visible, understandable and searchable regardless of weather or time. The ambition needs restraint as much as scale. Defense, commodity trading, disaster response and agriculture are not places where a plausible picture is sufficient. A model must tell the truth about a world that is inconveniently real.
The value of an uncooperative sky
Founders are often advised to look for pain. Yadav found his in a workflow, then followed it through layers of software until it met a cloud. That is the more interesting habit in his story. The company did not cling to the first interface it built. It kept the original frustration and allowed the answer to evolve.
There are also useful limits to the romance. AxionOrbital is young. Its benchmark numbers and product claims will be tested by researchers, customers and the inconvenient variety of the planet itself. The company has moved from India to a San Francisco headquarters, joined IIMA Ventures' Mavericks cohort and Y Combinator, and gathered a small team around Yadav and co-founder and CTO Atharva Peshkar. The work now has to outlast the launch posts.
Yet the founder's through-line is already visible. A student classifies crops from orbit. An engineer works with machine learning at India's space agency. An entrepreneur gets tired of repetitive geospatial tasks. A researcher studies how models should adapt to heterogeneous terrain. Then a CEO builds software for the hours when the familiar image disappears.
There is a modest management lesson inside that sequence. Expertise did not arrive as one immaculate credential. It accumulated as adjacent encounters with the same stubborn object: the Earth as data. Agriculture supplied the use case. Satellite imagery supplied the medium. Reinforcement learning and computer vision supplied tools. An earlier company supplied frustration. IIM Ahmedabad and two accelerator communities supplied people willing to challenge the plan. Yadav's job was to notice that these were not separate chapters. They were different angles on one problem.
The sky will remain rude. Clouds will continue arriving without checking the calendar. Night will fall on schedule. Yadav's wager is that visibility no longer has to leave with the light.