01 / System●
Pipeline
Photon → pixel → meaning
Detection, denoising, deblurring, demosaicing, color reconstruction, and semantic priors collapse into a single end-to-end network. No legacy ISP. No green.
REC · LIVE · 24·06·2026
SECTOR / R&D · BUILD 0.42.1
LAT 37.7956° N
LON 122.3933° W
ILLUM 0.0021 LUX
A research company
We are solving
night vision
with deep learning.
For seventy years, seeing in the dark has meant strapping a glowing green tube to your face. Tarsir Vision is replacing it with deep learning — models trained to turn a single photon into useful sight.
Founded
Connecticut · 2025
Stage
Seed
Discipline
Deep learning
Focus
Night vision
Deep learning for low-light vision◢Sub-lux color reconstruction◢Founded in Connecticut · 2025◢Currently seed funded◢Research-led◢Replacing the green tube◢Deep learning for low-light vision◢Sub-lux color reconstruction◢Founded in Connecticut · 2025◢Currently seed funded◢Research-led◢Replacing the green tube◢Deep learning for low-light vision◢Sub-lux color reconstruction◢Founded in Connecticut · 2025◢Currently seed funded◢Research-led◢Replacing the green tube◢Deep learning for low-light vision◢Sub-lux color reconstruction◢Founded in Connecticut · 2025◢Currently seed funded◢Research-led◢Replacing the green tube◢
01Thesis
The eye is a model. So is the camera you're about to build.
Conventional night vision is a brute-force amplifier. It collects whatever photons hit a tube, multiplies them with high voltage, and dumps the noisy result onto a phosphor screen. The hardware hasn't fundamentally changed since Vietnam.
We treat sight as inference. Our sensor counts individual photons across a wider spectrum than human vision; a model trained on millions of paired dark/light scenes reconstructs the world in real color, in real time, at one watt.
"The right architecture for a camera that sees in the dark is not a better tube. It's a model that knows what the world looks like."
— Tarsir Vision research notes
02Capabilities
One system.
Built end-to-end to see in the dark.
02 / Product●
Form
A camera, not a goggle
Twenty-eight grams. USB-C. Drop-in for drones, security cameras, robotics, automotive. The first SKU ships to design partners this winter.
FIELD SAMPLE · MOUNT TAM RIDGE · 0.0008 LUX
Same scene. Same moment.
Captured by the human eye, then by us.
UNAIDED · 0.0008 LXTARSIR · RECONSTRUCTED0.0005lux
Minimum illuminance for full color reconstruction
11ms
End-to-end latency from photon to pixel
5.8watts
Average draw of the production module
14PB
Paired training data captured across 6 continents
03Trajectory
From a dark room in Connecticut.
- 2025 · Q1FoundedTarsir Vision is started in Connecticut.
- 2025 · Q3First modelsEarly deep learning prototypes recover detail from photon-starved scenes.
- 2026 · Q1Seed fundingCurrently seed funded to accelerate research.
- 2026 · Q3Design partnersFirst evaluation partners come online.
04 / Request access
If you build things
that need to see
in the dark — talk to us.
We're working with a small number of design partners in defense, autonomy, robotics, and conservation.