The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
VisiHealth AI – Explainable Medical VQA
VisiHealth AI is a multimodal, explainable Medical Visual Question Answering system designed to analyze complex clinical radiology scans. Utilizing PyTorch, ResNet-50, and BioLinkBERT, the platform fuses text and spatial image patches through a real multi-token cross-attention mechanism to provide accurate clinical insights. It builds diagnostic trust through explainable AI by overlaying attention-map heatmaps on original scans and deriving human-readable rationales from a 4,444-triplet medical knowledge graph. Delivered via a scalable Next.js and Flask architecture, it demonstrates my capability in deep learning and healthcare AI integrations
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Related posts
One of our clients needed 70 ad video variants in a week, so we built an AI agent to do it in seven hours.
The build itself was a project. It required:
- Assembling a Claude agent connected to Remotion
- Building a knowledge base of the client's brand: brand guide, movement language, hook creation guidelines, past deliverables, briefing structure
- Teaching the agent to read the brief the way our team would
- Composing hook variants adapted for each aspect ratio (1x1, 9x16, 16x9)
- Assembling endcards and outputting files named and organized for QA
The first batch still needed heavy manual editing. The second, the agent assembled alone.
The third came out fully stitched (hook, body, CTA, disclaimer) needing only a review pass before shipping.
Each iteration meant testing what the agent got wrong, adjusting the skills, refining the knowledge base, and rerunning until the output matched the standard our team would ship.
When we sent the batch, the client feedback was: "The revised videos look great. We'll start using them immediately."
To show what this workflow can do, the video in this post was produced by this exact system.
If your business needs to ship ads at this volume, let's talk 👉 apta.agency/contact
70 ad variants in seven hours is wild. remotion?
Heron AI is an agent that reviews architectural models and flags building code problems. The hardest part of the brief was making that legible in the first few seconds.
We built the hero as a small interaction instead of a statement. The visitor drags across a black and white building sketch and violations surface where they drag, annotated with real IBC clauses. It puts the person in the agent's position for a moment.
Everything after that section can then talk in normal product language, because the core idea has already landed without needing to be described.
Really smart way to make a technical AI capability tangible—the drag interaction lets the user discover the value instead of reading a claim. The architectural visual language supports that clarity beautifully.
My own site, built to be edited in one place. Every section reads from a single data folder, so adding a project or a skill means editing one file, not the UI.
It ships with the SEO work done: sitemap, robots, schema.org structured data, social preview images generated at build time, and a plain-text summary for AI assistants. Verified in Google Search Console and Bing.
Stack: Next.js, TypeScript, Tailwind CSS, Framer Motion, deployed on Render.
The single data folder makes the site easy to update without opening the components. I'd add a build check that stops the deploy when a project is missing its title or social image.