Verifiable proof

Proof You Can Open in a Browser Tab

Working systems you can open right now, without talking to us. There are no client logos below and no testimonials — we will not publish proof we cannot substantiate. What we can put in front of you is software hjLabs.in built, deployed and keeps running in public. Open it, push on it, and judge the engineering yourself before you spend time on a call.

Live vision models, your GPU

Detection, segmentation and depth estimation running in your own browser

139open models
running live

Open any of these pages and the model downloads into your browser and runs on your own GPU through WebGPU — nothing is uploaded to a server. The catalogue behind them holds 139 open models across 22 task types, including OCR. It is the fastest way to see what current open vision models do and do not handle on your kind of image.

Regulated-data engineering

DICOM viewer with a full PS3.15 PHI de-identifier

652attributes
implemented

We transcribed DICOM PS3.15 Annex E Table E.1-1 — the Basic Application Level Confidentiality Profile — in full: 652 attributes, each carrying the standard’s own action code and VR. De-identification overwrites element values in place at their existing byte offsets, so the rest of the file, pixel data included, stays byte-identical to the original. It runs entirely in your browser; the study is never uploaded.

Hardware we actually build

We design and sell the machines, not only the models

5machine lines
sold from this site

hjLabs.in is a machine builder as well as an AI shop. The CNC router, selective soldering robot, wire cutter, spring coiler and IoT transformer winder below are our own products, on this same website, with real prices on them. The people writing the models also own the mechanics, the electronics and the firmware — very few AI consultancies can say that.

Why hjLabs.in — stated plainly

We are a small specialist team, not a large consultancy, and we are not going to dress up as one. Here is what that actually buys you.

You talk to the engineer who builds it

There is no account-manager layer between you and the person writing the code. The engineer who takes your scoping call is the engineer on your project, and stays reachable for the length of it.

Industrial hardware is our background, not a slide

hjLabs.in designs, manufactures and sells production machines — CNC routing, selective soldering, wire cutting, spring coiling, transformer winding. When a vision model has to survive a real factory line, cameras, lighting, PLCs and vibration are our day job, not a subcontractor’s.

Our work is inspectable before you pay us

Live tools you can open, open-source repositories you can read, and a standards implementation you can check line by line against the DICOM specification. You do not have to take a claim on trust when you can go and test it.

What we are not going to show you: client logos, testimonials or case-study numbers. We will not publish proof we cannot substantiate, and we would rather you find that out on this page than after signing something. For a focused engagement, the trade is straightforward — less brand-name comfort, more of the senior engineering time you are paying for.

What We Build with Computer Vision

Four use-case patterns we ship over and over — each with measurable ROI.

precision_manufacturing

Defect Detection & Visual QC

Real-time inspection of PCBs, weld seams, injection-moulded parts, textile rolls and packaged goods. We typically reach 96 to 99 percent recall on critical defects with a 1 to 3 percent false-positive rate. Models run on Jetson Orin at 30 to 120 FPS per camera. CV inspection pairs especially well with our in-house automated PCB soldering line — defective joints flagged at the camera close the loop back into the solder profile within seconds.

description

OCR & Document Processing

Invoice extraction, KYC document parsing, handwritten form digitisation, label and serial-number reading on the production line. Combines TrOCR, PaddleOCR and Donut-style document transformers with rule-based post-processing to hit 99+ percent field-level accuracy.

videocam

Surveillance & Security Analytics

Person/vehicle detection, intrusion zones, PPE compliance (hard-hat, vest, glove detection), licence-plate recognition, queue-length and dwell-time analytics. Edge-only options available for sites where footage cannot leave the premises.

flight

Agricultural & Drone Imagery

Crop-health analysis from multispectral drone imagery, weed/pest detection, fruit counting and yield estimation, tree-canopy mapping. We process orthomosaics in QGIS / OpenDroneMap pipelines and serve the predictions back as GeoJSON for farm-management software.

Computer Vision Tech Stack

The same tools we use in production, picked per workload — never “one stack fits all”.

Detection & Segmentation

YOLOv8, YOLOv10, RT-DETR, Detectron2, SAM 2, Mask R-CNN

Classical CV & Tooling

OpenCV, Roboflow, CVAT, Albumentations, Supervision

Edge Inference

NVIDIA Jetson Orin, Coral TPU, ONNX Runtime, TensorRT, OpenVINO

Cloud Vision Platforms

Vertex AI Vision, AWS Panorama, Azure Custom Vision, Roboflow Hosted

Engagement Process

A six-step path from first call to a model that earns its keep in production.

01

Data Audit

We look at sample images/videos, ambient lighting, camera angles and the existing label set. Output: a one-page feasibility note that says go, no-go, or fix-the-optics-first. If your existing dataset is low-resolution or noisy, try our free AI upscaler and denoiser on a handful of frames before the review — it often makes the optics call obvious.

02

Annotation Strategy

Pick label schema (bbox vs polygon vs mask), set inter-annotator agreement targets, and bootstrap with SAM 2 auto-segmentation on Roboflow to cut manual labeling 60 to 80 percent.

03

Model Selection

Benchmark three to five candidate architectures on your eval set. We optimise for the metric you actually care about (recall on critical defects, mAP at IoU 0.5, or character-error rate for OCR), not academic leaderboards.

04

Edge vs Cloud Decision

Latency, bandwidth, privacy, fleet size and total cost of ownership all go on a spreadsheet. We pick Jetson, Coral, on-prem GPU, or cloud — and tell you exactly why.

05

Evaluation

A held-out test set built from production conditions, plus a stress set of edge cases. Confusion matrix, per-class precision/recall, and a side-by-side video diff on real footage before sign-off.

06

Deploy & Monitor

CI/CD for models, drift dashboard (Evidently or Arize), one-click retrain, and an on-call SLA for the first 90 days. Customers retrain via alerts, not via panic.

Pricing & Engagement Tiers

Engagement-based, not per-token. Senior CV engineers, transparent scope.

Pilot CV

$3,000–$7,000

Single use case PoC, 3-4 weeks

  • Data audit + annotation plan
  • One model trained & benchmarked
  • Demo on real footage
  • Edge or cloud deployment guidance
  • 30-day support

Enterprise CV Platform

$60,000–$200,000

Multi-camera, multi-site, 16-24 weeks

  • 50+ cameras, multi-site rollout
  • Centralised model registry
  • RBAC, audit logs, SOC2-ready
  • Active-learning loop
  • Dedicated SRE on-call
  • 12 months support

Frequently Asked Questions

Six questions every engineering leader asks before greenlighting a CV project.

It depends on the task and how forgiving the deployment is. For a single-class defect detector with controlled lighting we have shipped working YOLOv8 models trained on as few as 600 to 1,200 annotated images per class, augmented heavily. For a multi-class detector that has to generalise across factories or seasons, we usually want 3,000 to 8,000 images per class. For OCR on bespoke document layouts, 200 to 500 fully annotated samples plus weak supervision from synthetic data is typical. We start every engagement with a data audit so you do not over-collect or under-collect, and we use SAM 2 plus Roboflow auto-label to cut annotation cost by 60 to 80 percent.

Edge wins when latency must be under 30 to 50 ms, when cameras run 24x7 and bandwidth is expensive, when the site has unreliable internet, or when video data cannot legally leave the premises. We default to Jetson Orin Nano/NX for high-FPS multi-camera workloads, Coral TPU for low-power single-camera installations, and on-prem GPU servers for plant-wide deployments. Cloud (AWS Panorama, Vertex AI Vision, Azure Custom Vision) wins when models retrain frequently, demand is bursty, or workload is centralised. A common pattern is hybrid: edge for real-time, cloud for retraining, audit and dashboards.

Optics first, model second. Most low-light failures we are asked to fix turn out to be lighting and camera choices, not model issues. We typically recommend dome or coaxial LED illumination for inspection, IR-cut switching cameras for outdoor surveillance, polarising filters for glossy surfaces, and global-shutter cameras for fast lines. On the model side, YOLOv10 with synthetic low-light augmentation plus a mild denoising preprocessor outperforms a fancy night-vision model in nine of ten cases. For genuine night work we use thermal fused with RGB. We include an optics review in every CV pilot.

For bounding-box annotation our blended cost is USD 0.04 to 0.12 per box. Polygon and pixel-mask annotation runs USD 0.18 to 0.45 per object. A pilot defect-detection dataset of 2,000 images with three classes typically costs USD 600 to 1,400 to fully annotate when we use Roboflow plus SAM 2 auto-segmentation with human review. For OCR we charge per page, not per word. We build a small auto-label pipeline in week one of every engagement so the cost curve flattens fast — by week three the model itself does 80 percent of the labeling.

Privacy is engineered into the pipeline, not added later. For EU/UK deployments we default to on-prem inference with no raw video leaving site, automatic face and licence-plate blurring at the edge, configurable retention windows, and audit logs. For workplace analytics we strip identity and only retain aggregate counts. For India we align with the DPDP Act 2023 and brief the customer DPO using our own DPDP penalty calculator (compliance.hjlabs.in). DPA, NDA and where required sub-processor agreements are signed before any footage is touched.

Plan monthly to quarterly retraining in year one, then quarterly to half-yearly. New product SKU, lighting change, camera replacement, a new failure mode, or measurable drift in the production confusion matrix all force a retrain. We ship every CV system with a drift-monitoring dashboard (Evidently or Arize) and a one-click retrain pipeline using Roboflow or Vertex AI Pipelines. Customers rarely retrain manually — alerts fire, a candidate is trained, the eval suite gates promotion.

Related Services & Industries

Explore the rest of the hjLabs.in AI/ML and automation portfolio.

Have a Computer Vision Idea? Let’s Stress-Test It in 30 Minutes.

Bring sample footage. We’ll tell you what data, optics and model class will actually work — and what will not.

calendar_todayBook a Free Scoping Call