Computer Vision That Works Outside the Lab
We build vision systems that hold up when the lighting is bad, the camera is dirty, the operator is bored, and the conveyor is moving faster than the demo. Real cameras, real edge devices, real factory floors — and models trained by engineers who've debugged a 92% mAP collapse caused by someone moving a fluorescent tube.
You don't need a demo on a clean image.
You need a model that works on the line.
Most computer vision projects look great in a Jupyter notebook on hand-picked images and fall apart the moment they meet a real camera. We build for the messy reality of vision in the wild: bad lighting, motion blur, occlusion, dust on the lens, the operator who tilted the bracket last Tuesday, and the model that needs to run on a $200 board, not a $2000 GPU.
Worked in the lab, broken on the floor
Trained on clean photos, deployed under fluorescent flicker, tilted brackets, and grease on the lens. Accuracy collapses in week one and nobody knows why.
The model only runs on a $5,000 GPU
Architecture chosen for benchmark glory, not deployment. Won't fit on the Jetson, won't hit the latency budget, can't be batched. The unit economics never work.
Labeled data is a mess
Three different annotators, no agreement on edge cases, half the boxes are wrong, and the test set leaked into training. The model is learning the labelers' bad habits.
What You Actually Get
No vague deliverables. Here's exactly what lands in your hands.
A vision pipeline running on real hardware
Camera capture → preprocessing → inference → post-processing → action. Deployed on the device you'll actually use, not just a notebook.
A clean, versioned dataset
Annotation guidelines, inter-annotator agreement, audited labels, train/val/test splits with no leakage. Stored where you can rebuild it.
Field metrics, not benchmark vanity
Precision and recall by lighting, by camera, by shift, by SKU. Drift dashboards on the actual deployment, not the validation set.
A model card and an ops runbook
Known failure modes, lighting requirements, recalibration steps, the operator's one-page guide, and the on-call playbook for when accuracy drops.
A Real Computer Vision Team
Shipping CV well takes more than a model and a webcam. Six roles you get on every Pillai Infotech vision build.
Vision Research Lead
YOLO, DETR, SAM, ViT, ConvNeXt — knows what's state of the art and ignores it when a smaller model wins. Trained models that run in 20ms on a phone CPU.
Camera & Optics Engineer
Picks the sensor, lens, focal length, exposure, and lighting. Knows that 80% of CV problems are camera problems and tells you so before you build the model.
Edge Inference Engineer
TensorRT, OpenVINO, CoreML, NNAPI, quantization, pruning, distillation. Squeezes a 200ms cloud model into a 30ms edge one without losing the accuracy that matters.
Data & Annotation Lead
Owns the labeling pipeline, guidelines, QA, active learning loop. Catches the edge cases the in-house team has been silently disagreeing on for months.
Deployment & Integration Engineer
GStreamer, RTSP, MQTT, OPC-UA, PLCs, the IT team that won't open a port. The engineer who actually gets the model talking to your existing system.
Field Reliability Engineer
Drift monitoring, recalibration playbooks, lighting checks, lens-cleaning schedules. Treats the deployment as a system, not a one-time install.
You See Everything. In Real Time.
Every Pillai Infotech project comes with a dedicated client dashboard. Kanban boards, live logs, test results, meeting notes — it's all visible the moment it happens. No status-report theatre, no "we'll get back to you", no surprises at the demo. You work with us like you work with your own team.
Kanban Board, Live
Every epic, every story, every task — visible on your dashboard. Drag, comment, reprioritize. It's the same board our team works from.
Documented Everything
Every decision, spec, API contract, and architecture diagram lives in the dashboard. Searchable, versioned, linked to the tasks they shaped.
Live Logs & Test Results
Build logs, deployment logs, test suite results — streamed to your dashboard the moment they run. You never have to ask "did the build pass?"
Meetings → Tasks, Automatically
Every meeting is recorded, transcribed, and every action point is auto-converted into a tracked task assigned to the right person. Nothing gets lost between calls.
Sprint Burndown & Velocity
See exactly how much work is done, how much remains, and our velocity over time. If a sprint is slipping, you see it the same moment we do.
Comment, Approve, Decide — In-Place
Comment on any task, approve designs, sign off on specs, and raise blockers directly in the dashboard. Everything tied to the work, not buried in email threads.
Vision Systems We Know How to Ship
We pick the architecture and the hardware to match the environment, not the other way round.
🏭 Industrial inspection & QC
Defect detection, dimensional checks, presence/absence, surface anomaly. Tuned for the false-reject and false-accept costs your line actually has.
📦 Logistics & warehouse vision
Barcode + OCR, dimensioning, parcel sortation, dock-door detection, damage capture. Built for low light, high speed, and operators who don't care about your model.
🛡️ Safety & compliance monitoring
PPE detection, restricted-zone intrusion, ergonomic risk, fall detection. Designed with privacy and consent baked in, not bolted on.
🛒 Retail & shelf intelligence
Planogram compliance, out-of-stock, price-tag OCR, queue analytics. Edge devices in-store, not mountains of video shipped to the cloud.
🏥 Medical & life-science imaging
Slide analysis, segmentation, measurement, triage support. Audit trails, regulatory documentation, and the explainability paperwork to back it up.
🚗 Mobility & ADAS-adjacent
Lane, sign, obstacle, license plate, vehicle re-identification. Multi-camera fusion, calibration, and the boring synchronization work that makes it actually run.
The CV Stack We Use
Boring tools where they win. Cutting-edge where they earn it.
Models
Edge Inference
Capture & Pipeline
Hardware
A Six-Stage CV Delivery Process
Built around the reality that the camera, not the model, decides whether this works.
Site Visit & Camera Audit
We come to the floor (or the store, or the line). Lighting, mounting, cabling, network, hardware budget. No models discussed until we see the environment.
Data Collection & Labeling
Real images from real cameras, in real conditions, across shifts. Annotation guidelines locked, inter-annotator agreement measured, edge cases documented.
Model Build & Offline Eval
Iterate on a holdout that mirrors deployment. Slice metrics by lighting, camera, time of day. No cherry-picked test images.
Edge Optimization
Quantize, prune, distill, benchmark on the actual target hardware. Latency, throughput, power, thermal — all measured before sign-off.
Pilot Deployment
One line, one store, one site. Run alongside the existing process for two weeks. Field metrics, operator feedback, calibration playbook.
Rollout & Monitor
Phased fleet rollout, drift dashboards, recalibration cadence, on-call runbook. Weekly review for the first 90 days, then handed off.
Three Ways to Engage
Vision projects don't fit one shape. Pick the one that matches your stage.
CV Feasibility Sprint
Two-week engagement: site visit, camera audit, data sample, baseline model. We tell you honestly whether vision is the right tool — and what the camera setup needs to look like.
- Site + camera audit
- Baseline + sample model
- Honest go / no-go in writing
Fixed-Scope CV Build
End-to-end vision system from data collection to edge deployment, with monitoring, calibration playbook, and post-launch warranty.
- Fixed scope, fixed price
- Typical: 12–24 weeks
- 60-day post-launch warranty
Embedded CV Squad
A dedicated CV + data + edge squad working alongside your team on a continuous fleet rollout.
- CV + Data + Edge + PM
- Monthly retainer, scale up/down
- Best for: multi-site rollouts
Honest Answers to CV Reality Questions
The questions every smart buyer asks before signing. Here's what we tell them.
Can you use our existing CCTV cameras?
Sometimes. CCTV is built for humans watching screens, not for models — low frame rates, heavy compression, fisheye lenses, awful low light. We'll audit what you have and tell you honestly which cameras can be reused, which need to move, and which need to be replaced. Often a $200 camera in the right place beats a $2000 one in the wrong one.
Cloud or edge inference?
Edge wins when latency matters, bandwidth is expensive, privacy is non-negotiable, or the network is flaky. Cloud wins when you have many cameras feeding centralized analytics and the latency budget is generous. Most real deployments are hybrid: edge for the hot path, cloud for retraining and analytics. We benchmark both for your case.
How much labeled data do we need?
Less than you think if the problem is well-scoped and pre-trained models cover most of it. More than you hope if you're detecting rare defects or new classes. We start with a sample, measure data efficiency, and use active learning to label only what actually moves the metric.
What about privacy and consent?
Designed in from day one. On-device processing where possible, face/license-plate blurring at capture, retention limits, consent signage, GDPR / DPDP / state-level rules. We'll file the paperwork and brief your DPO before the first camera goes live.
Can you handle bad lighting and dirty lenses?
Some of it, with the right augmentation, exposure control, and HDR. The rest is a hardware problem — better lighting, lens hoods, cleaning schedules. We'll tell you which is which, and we won't pretend a model can fix a fundamentally broken capture setup.
How do you measure accuracy in production?
Two paths. Sampled human review (a small percentage of predictions audited daily) plus passive drift monitoring on the input distribution. When ground truth arrives later (e.g. defect found downstream), we feed it back. Accuracy reports are sliced by camera, lighting, shift, and SKU — not one global number.
What hardware do you recommend?
Depends on the latency budget, model size, and where the camera lives. Jetson Orin Nano for most edge work. Coral TPU for ultra-low-power. Industrial PC for multi-camera. iPhone or Android when the camera is a phone. We pick after the camera audit, not before.
Who owns the model and the data?
You do. Model in your registry. Training images in your storage. Annotation tooling in your account. If we walked away tomorrow, your next CV team could retrain from a single command.
What if the lighting changes seasonally?
It will. Sun angles, daylight savings, store renovations, factory shift to LED. We bake recalibration into the runbook and use drift detection to flag it before accuracy collapses. Retainer clients get scheduled recalibrations included.
Can you sign an NDA before we share footage?
Always. NDA before the first call. Footage and model assets stay under your control. We're happy to work inside your VPC or on your hardware if compliance requires.