Computer vision or full-stack.Most often, both.

I build production computer vision models, ship full-stack web products, and join the two when a project needs the model and the interface to work as one.

View workDownload CV
Experienced in
PyTorchYOLOv11xFastAPINext.jsSupabaseBoT-SORTFAISSOSNetTailwindOpenCVMediaPipePyTorchYOLOv11xFastAPINext.jsSupabaseBoT-SORTFAISSOSNetTailwindOpenCVMediaPipe

Worked with / Interned at

Neural LinesPeriegesis

Real-time CV pipelines

Model-to-product delivery

LLM-integrated builds

Case studies

Projects that run in the real world.

Shipped computer vision systems and full-stack products.

Flagship case study - retail AI
In production

SecureVision turns live CCTV into real-time loss-prevention signals.

A shipped multi-camera pipeline from detection to ReID search and an operator dashboard.

YOLOv11x-poseBoT-SORTOSNet ReIDFAISSFastAPINext.jsSupabase
View project details

Problem

Cameras existed, but no reliable way to spot suspicious behavior before manual review.

Approach

An 11-stage pipeline: pose detection, BoT-SORT tracking, OSNet ReID, FAISS retrieval, and a Next.js dashboard.

Result

Deployed on live CCTV with GPU inference and sub-10ms ReID search.

  • Camera ingestion with frame sampling.
  • Pose-aware detection and multi-object tracking.
  • ReID embeddings, FAISS indexing, cross-camera lookup.
  • Behavior scoring on an operator dashboard.
View on GitHubWatch demo

Demo

SecureVision demo

Accessibility - hands-free control
In progress

Head-pose cursor control for hands-free computer access.

A lightweight HCI prototype mapping face pose to cursor movement.

MediaPipeOpenCVPython
View project details

Problem

Hands-free cursor tools often need extra hardware or brittle calibration.

Approach

MediaPipe Face Mesh landmarks mapped to yaw and pitch, with eyebrow-raise clicks.

Result

A calibration-light prototype built for commodity webcams.

  • Face mesh landmark extraction from webcam frames.
  • Yaw and pitch mapped to cursor velocity.
  • Eyebrow-raise gesture thresholding for clicks.
  • Smoothing and dead-zone logic against accidental movement.
View on GitHub
Health tech - full-stack AI productLLM-integrated
In progress

MediBuddy helps people understand health documents without losing context.

A full-stack product that turns uploaded health documents into a guided assistant.

Next.jsExpoNestJSSupabaseGemini APITurborepo
View project details

Problem

Health records are hard to search or explain, and scattered across systems.

Approach

Document intake, authenticated data flows, and Gemini-powered assistant responses.

Result

The full loop, frontend to LLM, working as one system.

  • Next.js web and Expo mobile flow for document access.
  • NestJS API layer for auth-aware workflows.
  • Supabase storage and database for user data.
  • Gemini API for document-aware assistant responses.
Open live site

Live site

MediBuddy live site hero reading Understand your health with an upload report call to action.

medi-buddy-lovat.vercel.app

Premium commerce - redesign
In progress

Caffeine Dealer redesign turns a coffee storefront into a sharper buying flow.

A premium e-commerce redesign built for high-intent browsing and product storytelling.

Next.jsGSAPTailwindTurborepo
View project details

Problem

The storefront needed stronger hierarchy and a smoother path to purchase.

Approach

Rebuilt with a Next.js frontend, GSAP motion, and product-led sections.

Result

A more distinctive surface that still keeps purchase easy to find.

  • Product-led structure with clearer conversion hierarchy.
  • GSAP motion for purposeful transitions and reveals.
  • Responsive layouts tuned for browsing and comparison.
  • Reusable patterns for future commerce pages.
Open live site

Live site

Caffeine Dealer live site hero showing espresso equipment and commerce calls to action.

caffeine-dealer.vercel.app

AI product - audit workflowLLM-integrated
Live

Audit Lens AI turns rough review work into a focused audit flow.

A deployed AI surface for reviewing inputs and extracting signal fast.

Next.jsGemini APITailwind
View project details

Problem

Review work gets scattered across documents and manual checks.

Approach

A focused review flow with AI assistance built in.

Result

A live demo of full-stack AI product thinking beyond computer vision.

  • Product flow built around review and next-step clarity.
  • LLM-assisted experience in a focused web interface.
  • Responsive frontend for demos and stakeholder walkthroughs.
  • Deployed live on Vercel.
Open live site

Live site

Audit Lens AI live site hero for receipt intelligence and finance workflows.

audit-lens-ai.vercel.app

Hire for a result, not a vague skillset.

CV model -> production API

Turn a trained CV model into a FastAPI service, ready to deploy.

Detection, pose, tracking, ReID, medical imaging prototypes

Starting at: let's scope itScope this

Full-stack + AI feature build

Ship the product surface around the model: UI, APIs, auth, and workflows.

Dashboards, review tools, internal AI copilots, MVP features

MVP to shipped featureScope this

Realtime vision pipeline audit

Review latency, ingestion, tracking drift, and inference so it survives real use.

Existing pipelines that need production hardening

Fixed-scope reviewScope this

From a YOLO experiment to production AI.

I started with a basic YOLO object detection model, partly out of curiosity and partly through coursework. It was the kind of small experiment that begins as a notebook and then refuses to stay there.

The turning point was realizing the model pointed at a real problem: retail loss prevention. That became SecureVision, my Final Year Project, designed as an 11-stage pipeline across YOLOv11x-pose, BoT-SORT, OSNet ReID, FAISS, PoseTransformer, and BehaviorAnalyzer, tested against live CCTV footage instead of clean demo clips.

Alongside the computer vision work, I have been shipping full-stack products end to end with Next.js, Expo, NestJS, and Supabase. Some of that work includes LLM-powered products, like a Gemini API health document assistant. I do not just train models. I build the product around them.

Right now I am focused on freelance work where AI has to survive real users, real latency, and real handoff. I am especially interested in computer vision systems, model-to-product builds, and full-stack AI features that need to ship cleanly.

First YOLO model

My first object detection model.

SecureVision FYP

Retail loss prevention, from detection to operator dashboard.

Full-stack + LLM products

End-to-end products with LLM features, not just models.

Computer vision & full-stack engineer

Freelancing on production vision models and full-stack products, together or standalone.

Computer Vision

YOLOv11x-posePose estimationReIDBoT-SORTPyTorchFastAPICUDA

Full-Stack + AI Products

Next.jsExpoNestJSSupabaseGemini APIPrisma

Notes from building real-time vision systems.

Technical writeups are being folded into the case studies first.

The details live inside the case studies for now.

Read breakdowns

Send the rough model. I'll ship the working version.

Computer vision meets full-stack product work. Send the messy state and I'll reply with next steps.

Email meUpwork
Responds within 24hPKT timezoneScope before build

What happens after you email

  1. Scope

    Send the repo, model, or rough idea.

  2. Build

    I connect the model, backend, and interface into one path.

  3. Ship

    You get a working handoff and a clear next step.