Computer vision, built to survive production
Every engagement here ends in the same place: a system that runs without someone watching it. I'm Muhammad Rizwan Munawar, a computer vision engineer and a top contributor to the Ultralytics YOLO project, and the tutorials, benchmarks and open-source tools on this site are the same engineering you would be hiring.
Everything I build
The full range, not a shortlist. If your problem is on here, it is one I have shipped before; if it is not, say so anyway and you will get a straight answer on whether it is solvable.
Object detection
Detection models trained on your footage and your classes, tuned to hold up on busy scenes rather than only on the test set.
You get: A trained .pt checkpoint, per-class precision and recall on your own holdout set, and an inference script that runs on a video file or an RTSP stream.
Object tracking
Multi-object tracking that keeps an identity across frames, occlusions and re-entries, so counts and paths stay stable over long clips.
You get: A tracker configuration tuned on your footage, validated against a manual count, with stable IDs across occlusions and a CSV of tracks per object.
Instance segmentation
Per-object masks instead of boxes, for measuring area and shape, cutting objects out cleanly, or redacting only what the object covers.
You get: Per-object masks with area and boundary output, exportable as PNG masks or COCO polygons for whatever consumes them downstream.
Image classification
Whole-image labelling at scale, by category, condition or defect type, including the long tail of classes that are rare in your data.
You get: A fine-tuned classifier, a confusion matrix over your real class distribution, and the confidence threshold that hits your acceptable error rate.
Pose estimation
Skeleton keypoints per person, for posture, ergonomics, rep counting, safety rules, or anything that depends on how a body is arranged.
You get: Keypoints per person per frame, plus the rule layer on top: angles, distances and the events you actually want flagged.
Depth estimation
Monocular depth from an ordinary camera, so distance, separation and rough 3D layout come out of hardware you already have.
You get: Relative or metric depth maps from your existing camera, with the calibration step that turns them into distances you can act on.
OCR & document AI
Pulling text, fields and tables out of invoices, IDs and forms, including scans and photographs that were never laid out for a parser.
You get: Structured JSON per document, field by field, with a confidence score on each and a review queue for the ones that fall below it.
Video analytics
Turning camera feeds into numbers: line crossings, zone occupancy, dwell time and flow, written to your database, dashboard or API.
You get: Counts, dwell times and zone events written to your database or API on a schedule, with a dashboard or a plain endpoint, whichever you already run.
Quality inspection
Defect detection on the line, for scratches, misalignment and missing parts, built around how expensive a miss is against a false alarm.
You get: A defect model tuned to your cost of a miss versus a false alarm, with the threshold set deliberately rather than left at 0.5.
Anomaly detection
Spotting the events you cannot collect enough examples of, by modelling what normal looks like and flagging what departs from it.
You get: A model of normal built from your own footage, an alert stream for what departs from it, and a tuning loop for the false-positive rate.
Data labelling & audit
Building and auditing the dataset: labelling policy, the hard cases first, and finding the broken images, bad boxes and split leakage.
You get: A labelling policy document, a cleaned and split dataset, and a report naming the broken images, mislabelled boxes and leakage found.
Model training & tuning
Training and evaluation with per-class metrics and the actual failure cases surfaced, so accuracy work targets what is really wrong.
You get: Training runs with the experiment log kept, per-class metrics, and the failure cases pulled out as images so the next iteration targets something real.
Edge deployment
Getting the model onto the device: NVIDIA Jetson, mobile and embedded targets, quantised and exported to hit a real latency budget.
You get: The model quantised and exported for your device, benchmarked on that hardware, with the measured latency and the fallback when it is exceeded.
Cloud & on-prem deployment
Serving the model where your data is allowed to live, on your cloud or your own servers, behind an API your systems can call.
You get: A containerised service behind an API your systems can call, with health checks, logging and a documented rollback.
Monitoring & retraining
Watching accuracy after launch, catching drift as cameras and conditions change, and retraining on the data the system has since seen.
You get: Drift metrics on live traffic, alerting when accuracy moves, and a retraining pipeline you can run without me.
What clients say
Every review is verified on Upwork, from clients who paid for the work described above.
6 🌟 reviews on- $2,620.80
Excellent to work with and exceeded all expectations in answering questions related to various courses. One of the standout qualities of Muhammed was their commitment to going above and beyond. They didn't just provide brief, surface-level answers. Instead, they took the time to provide in-depth responses that demonstrated their expertise and dedication to the job
VerifiedAugmented AIComputer vision project - $1700 / 55$/hr
Muhammad did an excellent job on my project. He's very knowledgeable, intelligent, and highly capable. I hope to work with Muhammad again in the future.
VerifiedJoe GratzDeep learning project - $600
Even though he had to work till midnight, Rizwan always worked diligently and attempted to solve difficulties. Rizwan is the best freelancer with whom I've ever worked in Upwork.
VerifiedYanuar BramantyoComputer vision project - $317.50 / 35$/hr
Good experience, was available when needed and on time on our appointment. explained well and teached well.
VerifiedDarmandran SivapathasundramComputer vision project - $270
Work was completed promptly and with great care. He's willing to go the extra mile and was willing to adapt to changes. Great job!
VerifiedAnn DelaneyComputer vision project - $400
Muhammad was very professional, helpful, and always available throughout the project. He provided valuable support and contributed significantly to the development of our solution. It was a great experience working with him, and I would definitely recommend him for similar projects. I hope we have the opportunity to work together again in the future.
VerifiedAndre VasquesComputer vision project
Before you book a call
What does Rizwan AI do?
Rizwan AI builds computer vision and deep learning systems that turn images and video into automatic decisions: detecting and counting objects, tracking people and vehicles, reading documents, and inspecting for defects. The work spans feasibility assessment through to a system running in production on cloud, on-prem or edge hardware.
How do engagements usually start?
With a short call, and then a look at your own footage or documents. Nothing can be promised on accuracy or cost before seeing real data, so that comes first. If the honest read is that the project will not work, you get that answer rather than a proposal.
Do you work with companies outside your region?
Yes. The work is remote and clients are worldwide. Camera footage, documents and deployment targets travel fine; where data cannot leave a network, the system is built to run on your own hardware instead.
Can you take over a project that has already stalled?
Usually yes. The first step is diagnosing whether the ceiling is the model, the data, or the problem definition. Those need very different fixes, and a stalled prototype often means the wrong one was assumed.