Average. The more I learn and grow, the more I realize how average we all are. It’s the hype that’s been created and fed to us — the narrative about ourselves, or rather the manipulation that has been happening for centuries. It’s built entire generations of imposters who present themselves as exclusive and brilliant, but deep down, everyone is just trying to survive systems we don’t really have control over.
Technically, we can do anything. We can grow in any direction. But we’re still bound by our mental capacity and the impaired attention of modern life — shortened focus spans, overstimulating environments, constant propaganda, subtle biases embedded in cunning systems.
So yes, we’re all Jack of all trades, master of none… and often stuck in all, rather than mastering even one.
I’m trying to practice my free will more consciously now — learning what I want, doing what I want, and saying what I actually want to say.
Lately, I’ve been focused and putting effort into what matters to me, and I’m seeing small wins along the way.
One of them is obviously tied to my ambition. What can I say? I’m an ambitious girl. I’m trying to live many lives in this one life.
And I’ve started understanding bioscience more deeply — it fascinates me more than I expected. It excites me. I can see the real impact of my work, and it finally feels worth it.
Yes, I know big pharma makes more money from what I do than I ever will — but the flip side is that I can reach far more people through a big system than I ever could alone.
Over the past few months, I’ve been building something I’m genuinely proud of — a Health Data Red-Team Protocol for Multimodal Imaging Pipelines. It started as a curiosity, but it quickly turned into a full research-grade system that stress-tests AI models used in ophthalmology.
Most clinical AI systems rely on a chaotic mix of fundus images, OCT volumes, ICD-10 histories, demographics, and visit timelines.
In reality, this data is messy — missing metadata, low-quality scans, laterality conflicts, contradictory ICD codes, wrong medications, or imaging that doesn’t match the chart at all.
That’s exactly where models hallucinate the most
— and where patient safety gets compromised.
So I built an end-to-end protocol that deliberately corrupts multimodal clinical data, feeds it into LLMs/multimodal models, and observes how safely they fail.
What the Project Contains
1. A full Python implementation
health_imaging_redteam.py simulates corruptions like:
- missing metadata
- incorrect laterality (OD/OS swaps)
- inconsistent ICD-10 vs imaging labels
- low-quality or truncated images
- conflict between medications, injections, and disease stage
It then evaluates whether the model:
- produces hallucinations
- contradicts clinical data
- suggests unsafe treatments
- overconfidently answers when it should abstain
2. A Research Proposal
The ResearchProposal.md file outlines:
- motivation
- problem framing
- methodology
- model evaluation metrics
- safety scoring
- future extensions (e.g., retinal biomarkers, OCT feature drift)
Project Link:
- A single-file consolidated version
Final Word
Working in bioscience changed me. It made me care about impact more than abstractions. Models aren’t just models — they affect real patients, real diagnoses, and real outcomes. If we don’t red-team them, they break in the worst possible ways.
So this project isn’t theoretical.
It’s a safety harness for AI in healthcare.
It’s the first time I felt like my technical skills — ML, infra, data engineering, imaging pipelines — converged into something meaningful.
I’m still average. We all are. But the work doesn’t have to be.