I build AI imaging systems that run in production.

Background replacement that leaves the object untouched, segmentation for medical scans, and the deployment work that keeps both fast and affordable at catalog scale. Below are four apps I built, two client projects with their numbers, and what clients wrote about the work.

Apps I built and run

Pick an app to watch it run. These are screen recordings of the live apps, captured on 24 October 2026. Both genaibg recordings are real runs.

genaibg.davincistitch.com
Open live site
Recordings

A product photo is cut out and placed in an AI-generated boutique scene, with the shadow kept. Recorded on a warm second run.

genaibg

Remove a photo’s background, then put in another photo, a colour, a blur or an AI-made scene. Or describe any edit in words.

The same workspace generates images from a prompt, repaints part of a photo, and handles voice and transcripts.

Speed
Recorded on a warm second run: background cut out and a new AI scene each in under 10 seconds. The first run after idle took about 50 seconds while the GPU worker started.
Models
FLUX.2 klein 4B (about 5 seconds) and 9B (about 10 seconds)
Output
Standard 1 MP, High 2 MP and Max 4 MP options, eight aspect ratios, up to five reference images
Replace
Paint over a region, or let the cut-out tool pick the object. PNG or JPG up to 2048 px.

Client work

Client names and private images are withheld. Every figure is as reported by the project’s own evaluation.

Automotive background replacement

A bespoke generative imaging pipeline for a client catalog of more than 100,000 vehicle images. It replaces the background while preserving the vehicle’s original appearance: body details, contours, windows, reflections and natural shadows.

No LoRA training workflow existed for the chosen model
I built custom tools and techniques to make the required training possible.
The client’s images needed preparation
We cleaned, categorised and annotated the supplied dataset as part of the workflow.
Glass, reflections and shadows are hard to composite
I developed techniques for window rendering and natural shadow integration, so the new background looks lit by the same light.
Quality, latency and cost compete
I analysed and tuned the workflow for sub-five-second processing, then oversaw evaluation, deployment and MLOps.
< 5 s
per image, in the tested environment. Latency varies with hardware and image size.
97%
reported background fidelity under the project’s internal evaluation, on 10,000 diverse automotive images.
100%
of original pixel dimensions preserved. This is not a claim of pixel-perfect output.
10 people
led across five engineers, two QA specialists, one backend, one frontend and one DevOps engineer.
12 months
from R&D through deployment and MLOps.

Brachial plexus segmentation from MRI

A research project with an NHS hospital in Birmingham, UK, to segment the brachial plexus and help identify whether an injury might be serious. The scans arrived as NIfTI 3D volumes and were processed as 2D slices.

Reliable ground truth
I engaged a clinical specialist to prepare and correct the reference annotations, and built specialised tools to make annotation practical for a complex structure.
Model selection
I ran multiple iterations and customised candidate approaches before settling on a strategy.
Lossy signal and difficult boundaries
I curated task-specific loss-function recipes, and combined generative AI with segmentation to isolate the nerve network.
91.27%
Intersection over Union, reported under the project’s evaluation protocol.

This is a research result. It is not an independently validated diagnostic tool, and it is not proof of clinical efficacy.

What I can do for you

High-fidelity background replacement

The object keeps its original pixel dimensions, and windows, reflections and shadows are handled so the new scene looks lit by the same light. See genaibg above and the automotive pipeline.

Large-scale AI deployment

From research prototype to a service that handles a 100,000-image catalog: evaluation, deployment and MLOps, with a team spanning QA, backend, frontend and DevOps.

Optimisation

Quality, latency and cost traded off on purpose. The automotive pipeline was tuned to under five seconds per image in its tested environment.

LoRA training

Custom tooling and techniques where no off-the-shelf LoRA workflow exists for the model you need.

Model fine-tuning

Segmentation and vision-language models adapted to your task, including loss design, ground-truth quality and evaluation. Examples: brachial plexus MRI segmentation, and a vision-language model for medical report analysis.

How I diagnose a failing model

  1. 1
    Read the evidence
    Inputs, failure examples, labels, results and constraints.
  2. 2
    Find the bottleneck
    Data, annotation, model, training, loss, post-processing or deployment.
  3. 3
    Run targeted experiments
    Controlled changes, not many variables at once.
  4. 4
    Verify the result
    Metrics plus a visual check on representative examples.
  5. 5
    Hand over for delivery
    Performance, cost, integration, QA and operations.

What clients wrote

Each one is a client’s own words about a job I delivered. The work closest to what I do now comes first.

“Ahmed has been incredibly helpful and excellent at following instructions. He completed the work to the highest standard and consistently aimed to improve the outcome, often suggesting alternative approaches. His knowledge of the segmentation task and proficiency with PyTorch were truly impressive … I would definitely work with Ahmed again in the future when needed and would definitely recommend.”
Segmentation task using PyTorch
September 2024, rated 5 out of 5
“Good analytical problem solving and core understanding of AI with proper reasoning.”
Fine-tuning a vision-language model for medical report analysis, phase 1
October 2025, rated 5 out of 5
“Thank you Ahmed for staying up all night to help me last minute with my work. Very helpful and knows what he is doing. Just hire him, trust me. You won’t be disappointed.”
Need help with image classification using PyTorch
Sobia K., UK. March 2023, rated 5 out of 5
“Ahmed was very professional, reliable and helpful throughout. The quality of service he provided was excellent, he did not disappoint.”
Algo & data structure
Lucien L., UK. May 2024, rated 5 out of 5
“I recently had a Python web scraping script developed by Ahmed, and overall, I’m pleased with the results. The script effectively scrapes the data as intended and does so with impressive speed. Ahmed was responsive throughout the project, addressing most of my concerns in a timely manner.”
Python script for web automation
Ashani H., Sri Lanka. August 2024, rated 5 out of 5
“Ahmed is a great partner to work with. I can highly recommend him, he is fast, always [available] and does the best.”
C/C++ developer for two small projects
Dino P., Austria. April 2024, rated 5 out of 5
“Ahmed converted 200 pages of non-OCR bank statements to CSV. He also kept in good touch through out and completed the assignment promptly. He is honest, open and straight-forward to deal with.”
Conversion of bank statements to CSV file
Jeremy W., UK. May 2023, rated 4 out of 5
“Ahmed’s done well in explaining matters and working on the job. I got to learn along the way which was my main goal. I would work with him again.”
Visual Studio, algorithms, and code
Caty B., US. February 2022, rated 5 out of 5

Send me the images, the scans or the catalog.

Tell me what you are trying to do. I will tell you honestly whether it is a good fit, and how I would approach it, before you pay for anything.