Working since
2021· İstanbul
An İstanbul research lab. We have been working on a single question since 2021: how do you train a model that can reason about the world the way the world actually presents itself — through video, lidar, audio and telemetry, in real time, with provenance you can defend? We are building toward our answer in public, in stages, starting now.
The next trillion tokens of useful intelligence are not on the web. They are streaming, in real time, from instruments that nobody has labelled yet.
A model that, given a short clip of an environment, can predict what will happen ten seconds from now, what an agent should do to achieve a goal, and what would have happened had the agent acted differently. Counterfactual, embodied, real-time.
This is the substrate, we believe, for the next decade of robotics, autonomy, climate science and industrial AI — domains where text alone is insufficient. We are not the only ones with this thesis. We may be among the most patient about it.
An encoder-only model trained on a small permissioned dataset on a single GPU. It is what we use to test data pipelines, evaluation harnesses and the parts of stratus we have written so far. Useful for nothing yet — by design.
Our first public release target. A 14-B distillation we plan to ship under a research licence with open weights once the data engine and evaluation infrastructure are ready. Currently in active development with funding contingencies.
The flagship model. A 70-B-class sparse mixture-of-experts trained on cloud-rented GPU capacity with our partner programmes. Realistic only with proper funding, a meaningful customer base and the infrastructure work between now and then.
One mobile multimodal capture rig running across İstanbul and three rural Anatolian provinces. RGB, lidar, IMU, GNSS and 8-channel audio, time-synced and on-device redacted before any frame leaves the rig. Honest scale: tens of terabytes, not petabytes.
The interesting part is not the volume — it is the provenance ledger we have written around it, which we treat as the most important code in the company.
RGB · stereo · 32-beam lidar · 8-channel audio · IMU · GNSS. Adding event cameras and 4D radar in 2027 once the calibration pipeline ships v2.
İstanbul, Konya plateau, two rural Anatolian provinces. Goal for 2027 is twelve cities across two climate zones — only if the design partner programme funds expansion.
Self-supervised pseudo-labelling and cross-modal consistency checks. Today: founder-only QA. Goal: a small expert annotation team funded through customer pilots, not headcount-burn.
Every byte ingested has a cryptographic provenance chain. On-device facial and license-plate redaction runs before transmission. Consent ledgers are stored on a tamper-evident log with row-level lineage to every model checkpoint we train. This is already in production for our own corpus, even though the corpus is small.
For every real trajectory we will render variants in a Gaussian-splatting twin of the scene to teach causal structure rather than mere correlation. The pipeline exists in prototype; full integration depends on cluster capacity we do not yet own.
Single-node H100 instances rented from cloud providers when we need them. Most of our work runs on a small on-premise development workstation. We are an active applicant to AWS Activate and the NVIDIA Inception programme, both of which provide compute credits to early-stage startups.
stratus · our training stack in progressOur internal training framework — early days, but starting to look like something. Designed from day one to elastically scale across heterogeneous rented GPU pools because we know we will never own the metal we train on at frontier scale. Open-sourcing planned for 2027 once we have something worth the README.
Closed-loop simulation, edge-case mining and behavioural prediction for fleet operators. We do not build vehicles — we believe the stacks that go into them deserve smarter eyes than a per-task CNN ensemble.
Continuous video understanding of factory floors, ports and energy infrastructure. Anomaly detection, predictive maintenance, OSHA-grade incident reconstruction. The vertical we expect to find our first paying design partner in.
Joint reasoning across satellite, drone and ground-station data — wildfire spread, urban flooding, methane plumes. We have a research relationship in early conversation with one civil-protection group.
Generalist manipulation and locomotion policies fine-tuned from BULUT for humanoid, quadruped and arm-robot OEMs. Our most ambitious vertical and the one that benefits most from a foundation-model backbone.
Latent dynamics, future-frame prediction, counterfactual rollouts. The mathematical core of what we want to build.
Shared representations across video, lidar, audio, IMU and text without paired supervision. Where we are spending most of our time today.
Mixture-of-experts routing, elastic restarts, fault-tolerant checkpointing across heterogeneous, rented GPU pools.
Generalist policies for humanoid, quadruped and arm-robot embodiments via fine-tuning from a foundation backbone.
Joint training on satellite, drone and ground-station data for nowcasting climate phenomena at sub-km resolution.
Byte-level audit logs, consent-aware training, model cards as cryptographic artefacts. Already running in production for our small corpus.
Three conversations are open right now: design partners (industrial / mobility / EO), AWS / NVIDIA / cloud credits, and engineers who like the patient-research-lab archetype. We answer mail in three languages within 48 hours.