Dhi-Technologies (DHI Technologies, Inc.)

Organization Card

Who we are

Dhi Technologies builds video analytics software that runs directly on Jetson class edge hardware, close to the camera, rather than shipping frames to a cloud model. Most edge vision systems answer every query with false confidence: a bare count, a bare detection, a bare alert, with no sense of when the system does not actually know. Dhi Labs builds the opposite: components that quantify their own uncertainty, ship a falsification ledger instead of a marketing claim, and say "I don't know" instead of guessing when they are outside their calibrated support.

Honesty, as a feature

Mechanism	What it means
Refusal gates	Components decline to answer when the input falls outside their calibrated support, instead of emitting a confident guess.
Calibrated confidence	Where a product reports an interval or a confidence score, the coverage of that interval against ground truth is itself measured and reported, including when it is imperfect.
Falsification ledgers	Predictive components log what they predicted, what actually happened, and whether the prediction was fulfilled or falsified, rather than only surfacing hits.
Stated limitations	Every dataset and demo states plainly whether its numbers are synthetic or real world, and what has not been checked yet. No claims of state of the art or foundational status, only what was actually measured.

Two more things follow from the same policy: no SOTA claims, no customer names, and no deployment claims (this org is a public research and engineering surface, not a sales page), and every benchmark below is synthetic first and disclosed as such, so every claim can be checked exactly while real world validation is still in progress.

Products and evidence

Six products, each with a gated benchmark dataset built from procedurally generated ground truth and a live static demo Space you can exercise in the browser.

Product	What it does	Demo	Dataset
Amodal Counting (A4)	Visibility corrected counting through crowds and occlusion: a calibrated interval instead of a bare point count.	Space	Dataset
Multicam Reasoning Memory (E1)	Cross camera identity linking with transit time priors and weeks scale bounded memory; every answer carries provenance, and it refuses to link when unsure.	Space	Dataset
Causal Predictive Alerting (E4)	Predicts an incident seconds before it happens from kinematic trajectories, and proves why via counterfactual replay and a fulfilled or falsified ledger.	Space	Dataset
Fixed Camera 3D (A3)	Turns an ordinary fixed camera into a metric 3D sensor by self calibrating from people walking through the scene: no GPU, no model weights.	Space	Dataset
Thermal Perception (A5)	A radiometric data engine and self supervised pretraining harness for thermal native perception, with CPU verifiable math ahead of any GPU pretraining run.	Space	Dataset
Prompt2Model (B1)	A language guided vision model factory: prompt to dataset to trained model to a calibrated, quantized, ONNX exported artifact, with an accuracy floor refusal gate.	Space	Dataset

Everything above, plus the blog dataset, is indexed in one place: the Dhi Labs collection. Browse all datasets or all Spaces directly.

Writing (the blog)

Four technical posts, each grounded in the committed repo numbers. Hugging Face has no public Posts or Articles API, so these live as versioned markdown in the blog dataset.

AI that refuses to guess: the thesis, calibration, refusal gates, provenance, and falsification ledgers as product features.
Precision first cross camera linking (E1): how a uniqueness guard lifted synthetic site precision from 0.918 to 1.0, and how weeks scale memory stays under 2 MB.
When the error bar is the product (A4): calibration honesty, the analytic interval undercovered (0.50 to 0.63 versus a claimed 0.90); conformal calibration widened it 6.7x to 14.5x.
Six products, one honesty thesis: a portfolio overview tying all six products to the honest by construction thesis.

Research status

Research papers describing these methods are in preparation and have not yet been published on arXiv or any other venue. The two whitepapers on the labs page are internal architecture write ups, not peer reviewed papers, and are labeled as such.

Code and evidence

The code itself is proprietary and closed source permanently. These gated datasets, benchmarks, and demo Spaces are the published evidence: closed source code, open evidence. Prompt2Model is a deliberate exception with one public, MIT licensed release: Prompt2Model, Language Guided Vision Model Factory. The GitHub org linked below is an identity link, not a browsable destination for the other five products.

Links

Website: dhi-tech.com
Labs (live page, two whitepapers): dhi-tech.com/labs
Collection (the hub): Dhi Labs, honest edge vision AI (all six Spaces, all datasets, and the blog in one place)
Spaces: amodal-counting, multicam-reasoning-memory, fixed-camera-3d, causal-predictive-alerting, thermal-perception, prompt2model
Datasets: browse all Dhi-Technologies datasets, or the table above
Blog: Dhi Labs technical posts
Code: proprietary, closed source permanently. These datasets, benchmarks, and demo Spaces are the published evidence; the code itself is not released. Prompt2Model is a deliberate exception with one public, MIT licensed release: Prompt2Model, Language Guided Vision Model Factory. The GitHub org github.com/DHI-Technologies-Inc is an identity link, not a browsable destination for the other five products.
Partnership or access inquiries: dhi-tech.com