Dhi Labs

Edge vision AI that
refuses to guess.

Honest AI for the physical world.

Calibrated intervals instead of point guesses. Provenance on every answer. Systems that grade their own predictions and refuse to ship what they can't verify. Six research systems, built and measured this way.

Runs on Jetson-class hardware in the ₹40,000 class, no data-center GPU required, and the platform underneath is engineered and hardened for national-scale transit and public-infrastructure operators, validated to production-grade edge requirements.

Honesty is the product, not a caveat

Every edge vision system eventually meets a scene it can't resolve: an occluded crowd, an ambiguous re-identification, a prediction that may or may not come true. Most systems paper over that with a confident-looking number. Dhi's systems are built to say "I don't know" in a specific, checkable way: a refusal gate that won't ship a degraded model, a calibrated interval instead of a bare estimate, a provenance trail on every answer, a ledger that grades a system's own alerts against what actually happened.

That's the differentiator we're building toward: not a bigger model, but a class of mechanisms (refusal, calibration, provenance, falsification) that make the system's confidence auditable instead of asserted. Below is each research system, with the honesty mechanism and one measured number per system, and a link to the evidence: the benchmark dataset and live demo for each. The system code itself is proprietary; see the footer note.

Refusal gates
Calibrated intervals
Provenance
Falsification ledgers

Six publicly demonstrated research systems

Each system is tested and benchmarked on synthetic ground truth today (see the footer note). The code is proprietary; the benchmark dataset, demo, and README linked from each card are the published evidence.

A4 · AMODAL-COUNTING

Visibility-corrected crowd counting

Corrects a raw detection count for what the camera can't see, instead of reporting the visible-only number as if it were the true one.

Calibrated intervals
Synthetic: MAE 3.73 → 2.46 (naive vs. corrected) at the hardest occlusion density tested, with 90 to 97.5% empirical coverage on its conformal intervals across density levels. Real data: on 99,481 validation images from a public crowd and occlusion benchmark, the learned amodal box reaches mean IoU 0.7677 versus 0.6768 for a deterministic identity baseline (a weak comparator that just assumes the visible box is the full extent), and visibility MAE 0.1096 versus 0.3177 · 61 tests.
A3 · FIXED-CAMERA-3D

Metric 3D from one existing camera

Ground-plane position, height, and separation from a single fixed camera, self-calibrated from the person boxes already passing through frame, no LiDAR, no stereo rig.

Measured, not asserted
Position RMSE 0.16 m (5 m height, 30° tilt) to 0.19 m (8 m, 45°) · 51 tests.
A5 · THERMAL-PERCEPTION

Thermal perception, sensor gap disclosed

A self-supervised recipe for thermal perception that is explicit about what's proven today versus what a GPU training run would still need to confirm.

Validated probe, not a fabricated benchmark
Two real Dhi Thermal (ViT-B MAE) runs now exist on different public thermal corpora (15,488 frames from a public thermal pedestrian dataset; 47,706 multi-source frames). A validated 4-way person-count probe (majority baseline 0.3237) shows both beat their own random-init control: 0.6136 accuracy against a 0.5379 control, and 0.5158 against a 0.4311 control. Neither clears our 200k-frame production floor, and the ranking between the two corpora is not resolved. Getting to a trustworthy probe meant catching a near-degenerate evaluation and a corpus-contamination bug first, written up in full here · 151 tests.
E4 · CAUSAL-PREDICTIVE-ALERTING

Alerts that grade their own promises

Causal, predictive alerting for early warning, paired with a ledger that checks every lead-time promise against what actually happened, after the fact.

Falsification ledger
200-scenario battery, graded against itself: 63 fulfilled · 87 falsified · 50 no-alert · 22 tests.
E1 · MULTICAM-REASONING-MEMORY

Cross-camera memory (“fleetmind”)

Links the same person or vehicle across cameras and answers where/when/who-was-there questions with a provenance trail on every answer, and refuses to link when it isn't confident.

Precision-first, provenance-carrying
Precision 1.0 (recall 0.377), zero wrong merges on the seeded demo; a simulated month of events stays under 2 MB · 47 tests.
B1 · PROMPT2MODEL

English spec → deployed TensorRT engine

A compiler for vision models: describe the detector you want in plain English, get a deployed, benchmarked TensorRT engine, or a documented refusal, not a silent degradation.

Refusal gate
The compression gate refuses to ship a compressed artifact that falls below its accuracy floor and keeps the uncompressed one instead. Applied to Dhi's own production edge models: the VideoMAE violence classifiers compress 70.5% (about 90MB to about 26MB each) at 0.9999+ output-fidelity agreement with their FP32 originals; the FaceLiVT face-embedding model compresses 6.5%. Conv-heavy detectors (YOLO, PPE, the SCRFD face detector) gain almost nothing, since this pass only quantizes MatMul/Gemm nodes, an honest negative we are publishing, not hiding. Net: roughly 124MB freed on a memory-constrained Jetson Orin · 128 tests.
E3 · EDGE-SCENE-GRAPHS

Neuro-symbolic scene graphs

Turns detector output into queryable relations plus a symbolic rule engine, with plain-language explanations, and a zero-dependency core so it runs on Jetson-class devices as-is.

Details → · Closed source, inquire →
A2 · CONTINUAL-OPEN-VOCAB

Continual open-vocabulary detection

Add a new alert class by text prompt or image crops, on the box, forward-only, no cloud retrain, no GPU training loop, no forgetting of prior classes.

Details → · Closed source, inquire →
A8 · FACE-FREE-REID

Face-free re-identification

Long-term identity across cameras from appearance, gait, and body shape and learned transit priors, deliberately with no face biometrics.

Closed source, inquire →

One general platform underneath

The research systems above plug into a single edge-analytics platform, not six separate stacks.

26canonical use cases across 11 detector engines, config-driven, not a single-purpose tool
1×8GBNVIDIA Jetson Orin Nano runs the whole pipeline, no data-center GPU
~710MBtotal memory for 6 cameras on one shared inference engine, down from ~3.4GB when each camera ran its own process
Transit-gradeengineered and hardened for national-scale transit and public-infrastructure operators, validated to production-grade edge requirements

The platform also carries a base-kitchen food-safety build for regulated food-service operators: on-floor object detection, contamination-zone dwell alerts, camera-health monitoring, and P1/P2/P3 alert triage, shipped as config-driven handlers per the operator's food-safety spec. Pest and PPE detection in that build are model-gated: the code path is real, the detector weights are not yet installed; we say so rather than claim a capability that isn't running.

Research track

A portfolio of applied-systems papers in preparation, built from the same code and numbers on this page, not a separate marketing narrative.

Reproducible benchmarks and open evidence are the norm for every paper we ship, not an appendix; the code behind them is proprietary.

Whitepapers

Blog

Shorter technical write-ups on the same research track, published as versioned markdown on Hugging Face.