AberTech
Computer Vision with Clarity
← Main
Welcome

Proof-of-concept demos, built to ship.

AberTech builds cloud-native computer-vision and geospatial pipelines. Explore the Projects tab for five open proof-of-concepts — from movement-aware CCTV archival with a Google Cloud deployable demo to UAV litter counting — each backed by a public repository.

It began as "Broad CoCo": a search for one standard unifying QC and pre-processing across vision projects. The approach owes more to 3D animation on low-end hardware than to textbooks — render only what the viewer needs, know your light sources, and treat memory as scarce, because it is.

About AberTech

You are already giants. AI should stand on your shoulders.

Named for a Welsh valley town that offered a moment of total clarity, AberTech is hired for one job: making computer vision serve the systems you already run. We don't need more data — we need better examinations of data. Every engagement starts with your metadata, strips out what never needed analysing — respecting your carbon footprint as much as your budget — and leaves the final 20% open-ended so your IT and security teams finish it on their terms. Proof-of-concepts run on mock data first, because a partner who can't mock your data doesn't understand your objectives. AI is not about replacement — it's about adding to a pre-existing system, without restructuring your business to meet a buzzword. And for environmental work, one rule is absolute: it makes no sense to analyse the environment by environmentally unsustainable means.

5 Open POCs
2 Cloud-Deployable Demos
3 Domains: Vision / Geo / Cloud
The AberTech Method
  1. Deduce what data is actually relevant — and what the raw mock data should be.
  2. Pre-compute whatever the metadata already gives you; estimate regions of interest mathematically before reaching for a CNN.
  3. Describe every vision task in plain photographic properties — brightness, texture, line composition — so sanity checks stay honest.
  4. Isolate control, independent and dependent variables between frames, noting what is stationary and what moves.
  5. Shrink the search space before any model runs — remove the regions where the answer cannot physically be, as in the offshore wind-farm detection model.
  6. Exploit material physics — noise signatures, reflectance, and the manufactured simplicity that separates synthetics from nature.
Get in Touch

Start a conversation.

Direct ikramchoudhury0@gmail.com Code github.com/Ikram-c

"AI is about prediction — how can you achieve that if you can't do the starting data entry process?"

— A lesson carried from British Airways
Select a project
Cloud-Optimised CCTV (cctv_zarr)
Search a week of footage, download only the minutes you need. Movement-aware OME-Zarr archival with chunk-level cloud query — Google Cloud deployable demo.
Cloud Dashcam POC
Keep only the frames worth keeping. Quality-gated, illumination-aware, coverage-chunked frame extraction from vehicle-mounted video.
HexCover
Turn any named area into an answerable map. H3 hexagon covering, EO scene availability, WEBKNOSSOS visualisation — retargets via YAML alone.
UAV Unsupervised Litter Counter
Count litter from drone imagery without labelling a dataset. Unsupervised counting via OpenDroneMap overlap pairing — separating synthetics from nature by shape, because nature has never been composed of simple objects. Counting is the start; the goal is understanding the human behaviour behind where litter lands.
Vision Toolkit
Skip a year of dataset plumbing. Eleven CV / geospatial / dataset codebases consolidated into one monorepo.

"We don't need more data, we need better examinations of data."

— AberTech ethos

"The world has plenty of coders — we need more scientists."

— AberTech