Full stack builds
Your whole website or app, designed, built and put online.
Frontend, backend, database, deployment, all of it in one pair of hands. Something that runs and keeps running, not a demo that falls over on the second click.
Applied AI + Full Stack / 5 Years Building / MS Robotics & AI, Arizona State
I build AI that works outside the notebook. Detection systems, language models, and the full stack that carries them to something people can actually use.
The gap
Plenty of things look brilliant in a notebook. Vision models that are perfect on clean validation images start guessing the moment the light goes bad. Assistants that demo beautifully go sideways the first time somebody asks the question nobody planned for. Sites that run fine for you fall over the week they finally get traffic. I work in that gap.
Five years of it, across production systems, research labs and contract projects. Most of the job is the unglamorous half: making something that works in principle work on real hardware in real conditions, fast enough that nobody complains. Language models tuned with human feedback so they answer the way one business needs rather than the way the internet would, detection and tracking that holds up on a warehouse floor at 6am, and the web and data plumbing that carries either one to somebody who can actually use it.
If that sounds like your project, a conversation is a good place to start. No finished brief needed, and no obligation either way. And if it needs one person who can go from raw data collection to a deployed endpoint without three handoffs, that is the job I take.

Hire me for
Keep scrolling →
Your whole website or app, designed, built and put online.
Frontend, backend, database, deployment, all of it in one pair of hands. Something that runs and keeps running, not a demo that falls over on the second click.
A chatbot or assistant that actually knows your business, not a generic one.
Retrieval over your own documents, assistants wired into your existing tools, model endpoints your app can call. Built into the product, not bolted on beside it.
Software that watches a camera and tells you what it sees.
Detection, tracking, counting, inspection. Multi-camera pipelines that hold up in a warehouse at 6am, not just on a clean validation set.
When an off-the-shelf model does not fit your problem, I build one that does.
Architecture, training loop, loss design, evaluation. From scratch or fine tuned from an existing model. I will tell you which one you actually need before you pay for the expensive option.
Turning the mess you already have into something a model can learn from.
Collection, annotation, labeling standards, cleaning, versioning. The unglamorous half that quietly decides whether your model is any good.
Making it run on real hardware, for real users, at a cost you can predict.
Jetson, CUDA and TensorRT when it has to run on the device, managed cloud hosting when it has to run for everyone. Quantized, profiled and measured before I call it done.
Live demos
A couple of working models, on their own page. Ask a document a question and watch it find the answer step by step, or point an object detector at a photo. Both run in your browser, so anything you load stays on your device.
Live demos
Pick one and give it something of your own if you like. Models download only when you ask for them and run on your device. The only thing that ever leaves is the retrieval demo's final question, along with the few passages it picked out.
A short story, about 3,400 words. Nothing has downloaded yet.
POSITION IN DOCUMENT
RANKED BY COSINE SIMILARITY
GROUNDED ANSWER
This is retrieval augmented generation, or RAG, the thing sitting behind most document chatbots. The embedding model runs in your browser. Your document is chunked and embedded on your device, and only the few passages that match your question are sent to be answered.
Nothing loaded yet. The model is about five megabytes and only downloads if you ask for it.
Runs entirely on your device. No image or video is ever uploaded. Object detection and segmentation available on request.
A 190 KB model I trained on 112,800 handwritten characters. It reads most things cleanly and will tell you when a stroke is genuinely ambiguous.
Learned filters, conv1
Nobody designed these. The network arrived at them on its own.
Draw a few characters with a gap between them, then read it.
Era two. A sequence model I trained on thirty handwriting fonts, on my own machine. It reads the whole word at once, so joined-up writing works, and it never cuts the image into characters. 11.9 MB where the other era is 190 KB. Reading in context costs.
Conv stack, 4 of 256 channels
BiLSTM state, 512 units256 forward filling left, 256 backward filling right
Per-frame guess, then the collapse
Write across the pad, joined or separated, then read it. It reads a line of writing, so wide works better than tall. First read downloads the model, 11.9 MB.
These are the smallest models that still show the capability, picked so they download in seconds and run on whatever device you happen to be holding. Purpose-built models outperform these.
How I work
I take pride in how quickly I move. Every tool that shortens a build without costing quality gets used, and I keep looking for better ones.
For you that means a shorter timeline and something real in front of you early, rather than a long quiet stretch before the first look.
Solid frameworks, automation, and whatever else keeps repetitive work off the critical path.
You see something running quickly, so you can react while changing direction is still easy.
The pace comes from good tooling and from having built this kind of system before, not from skipping the parts that matter.
No jargon
Most people arrive with a problem, not a spec. That is completely normal. Here is what those problems usually sound like, and what I would actually build for each one.
“I have hours of camera footage and no idea what is in it.”
A system that watches the video for you and reports what happened, where, and when. You get a searchable record instead of a hard drive.
“I want a chatbot that actually knows my business.”
An assistant connected to your own documents and data, so it answers from your material rather than making things up.
“We count or inspect things by hand and it is slow.”
A camera and a model that do the counting, flag anything that looks wrong, and keep a log you can check later.
“We have an idea but no website or app yet.”
The whole thing built and live. The pages people see, the database behind them, and the hosting so it stays up.
“Someone built us an AI feature and it does not work.”
I take it apart, find where it actually breaks, and tell you honestly whether it is worth fixing or rebuilding.
“We have plenty of data but it is a mess.”
Cleaned, labeled and organised properly, so the model you build on top of it is worth the money you spend training it.
Toolkit
The full list, in case you want to check something specific. If you would rather skip the jargon, the section above covers the same ground.
Track record
Every one of these came from a system that had to keep working after I stopped looking at it.
Sales growth on a product built around the detection architecture and camera integration I designed.
Social chatbot performance gain using RLHF and multi-GPU training.
Virtual therapy agent effectiveness gain in the same lab, from the same training stack.
ETL latency cut with PySpark, alongside wiring Unreal Engine 5 into PyTorch Profiler for healthcare and VR simulation.
Off the clock
Blues, hard rock and metal. Mostly the parts with too many notes in them. Years of playing to rooms taught me more about reading an audience than any deck ever did.
Simulator hours now, track time next. Same instinct as the engineering: find the limit, then find out what it costs to hold it there.
I also give time to NGOs when there is something worth building for them. Good engineering should not only be available to people who can pay for it.
Contact
Open to contract and freelance work, including long engagements. Some of the best things I have built ran a year or more with the same team. Tell me what you are trying to ship, in whatever words you have, and I will tell you straight whether I am the right person for it.
The quickest route is the chat in the corner. It reaches me directly, so leave your details there and I will pick it up.
Like this site? I built all of it, demos included. Ask the chat to set up a call and we can talk about yours.
Assistant
A transcript of this chat is emailed to Eshan so he can pick up where you left off instead of asking you to explain everything twice. Nothing else is stored, and it is not shared with anyone else. If you would rather not have that happen, use the email link on the page instead.