01
Restored frame
A clean, colour-true image — the only input the pipeline needs.
Computational Marine Imagery · ABYSS-1 uplink
Reveal the Hidden Depths of the Ocean
You are looking through the eye of an underwater rover, 2,400 metres down. The water has taken almost everything from this frame. Keep scrolling — we're going to give it back.
Chapter 06 · Abyssal Studio
A restored frame is only half the answer
Clarity is the input, not the product. Everything the vehicle sees goes into Abyssal Studio — our analysis workbench — where a single click turns pixels into a catalogued, confidence-scored, exportable survey.
0K+
Species classes
0%
Detection mAP
0ms
Detect latency
IOD
+ EUVP + UIEB
Chapter 07 · Reconstruction
From one frame to a world you can walk through
Monocular depth estimation lifts each restored frame off the image plane. Consecutive frames fuse into a georeferenced point cloud, then a textured mesh — a survey site you can open in Blender, QGIS, or your own viewer.
02
Depth field
Per-pixel range inferred with a marine-tuned monocular depth model.
03
Point cloud
Pixels lift into 3D and fuse across frames with DVL-aided odometry.
04
Textured mesh
Poisson surface reconstruction, then projected colour from the source frames.
.glb.obj.ply
.lazBlenderQGIS
Geo-referencedSediment classes
Chapter 08 · The platform
One suite, four stages, one vehicle
There is no model to pick and no mode to switch. The suite decides what a frame needs, runs it, and hands the next stage a better input than it had.
Stage 01
Restore
Colour-cast inversion, backscatter removal and 4× super-resolution, quantised to INT8 and running in the thruster power budget.
- UDnet unpaired restoration
- LU2net + MobileIE edge inference
- 4× super-resolution, 30 ms/frame
Stage 02
Perceive
Detection and classification over 15,000+ marine classes, plus geology and threat catalogues, every box confidence-scored and exportable.
- YOLO detector on IOD / EUVP / UIEB
- 98.5% mAP at 50 ms
- Species, geology and threat heads
Stage 03
Reconstruct
Frame-to-frame fusion into georeferenced point clouds and textured meshes, with sediment and structure classification baked in.
- Monocular depth + DVL odometry
- Poisson meshing, projected texture
- GIS and Blender-native export
Stage 04
Decide
Telemetry, quality scores and CLIP-validated retraining close the loop: every dive makes the next dive's models better.
- PSNR / SSIM / UIQM scoring
- CLIP validation gates training data
- Nightly retrain, OTA weight rollout
Raw feed
Restore
Detect
Reconstruct
Survey output
Chapter 09 · Deployment
Dual-use, by design
The same stack that helps a marine biologist count coral recruits helps a navy classify a contact. It runs on the vehicle, on a ship's server, or on a diver's headset.
Defence & security
Maritime surveillance, submarine and mine classification, coastal monitoring, unauthorised-diver detection.
Offshore & energy
Pipeline and riser inspection, rig structural surveys, anode and corrosion tracking, site clearance.
Research institutes
Reef health time-series, biodiversity cataloguing, climate monitoring, deep-sea and hydrothermal exploration.
Survey & dive operations
Live visibility enhancement for pilots and divers, wreck documentation, guided species identification for tour operators.
Chapter 10 · Mission & team
We are building the ocean's missing sense
Seventy percent of the planet is dark to us. Not because we lack cameras — because water destroys what cameras record. Underwater AI exists to give that sense back, cheaply enough that every vehicle can carry it.
Goal 01
Real-time, on the edge
Everything must run on the vehicle inside a 15 W budget. A model that needs a ship's server is a model that arrives too late.
Goal 02
Sensor-agnostic
GoPro, machine-vision rig, or a 4,000 m-rated ROV camera — the same weights work, with no per-camera calibration pass.
Goal 03
Evidence, not pictures
Every output carries confidence scores, quality metrics and provenance, so it stands up in a survey report or a mission log.
Goal 04
Built for the Indian Ocean first
Trained on the Indian Ocean Dataset alongside EUVP and UIEB, because the waters we know best are the ones least represented in public data.
The founders
Gautam Singh
Chief Executive Officer
Sets the vision and drives strategic partnerships across defence, energy and research.
Shuvam Banerji Seal
Chief Technology Officer
Leads model architecture, edge deployment and the engineering team behind the suite.
Youktik Sajjan
Chief Operating Officer
Orchestrates operations, partnerships and field deployments across dive programmes.
Aman Kumar
Chief Product Officer
Shapes the product roadmap, the Abyssal Studio experience and customer success.
MeitY · Government of India
Funding partner
Funding partner
Founded 2024 · India
Seed stage
Seed stage
Take it down there
Put Underwater AI on your vehicle
Tell us what you fly, where you dive, and what you need to see. We'll send a build tuned to your optics and your water.