Skip to content
View mghadia1's full-sized avatar
🎯
Focusing
🎯
Focusing

Block or report mghadia1

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
mghadia1/README.md
Mayank Ghadia — Perception, Planning, Honest Evidence

Portfolio LinkedIn Email



COMPUTER SCIENCE SENIOR  ·  MEDICAL AI  ·  ROBOTICS  ·  ML


Profile — A senior learning this properly


I'm a Computer Science senior at Arizona State University (GPA 3.63, graduating December 2027) focused on medical AI, robotics, machine learning, and deep learning.

I learn by implementing things from scratch first, then reading the theory, then explaining it back in my own words. A topic isn't finished until I can defend it.

What I build

Medical-image segmentation, predictive maintenance, retrieval systems, model adaptation, and robotics simulation.

How I work

Real or public data, held-out evaluation, reproducible pipelines, and write-ups that preserve what failed.

What I'm seeking

A 2027 internship in medical AI, robotics, computer vision, machine learning, or applied AI engineering.


Projects — I build in order and keep the failures in the write-up


These twelve projects form a deliberate sequence—not a pile of demos. Each one teaches what the next one needs: gradients before frameworks, an explainable vision baseline before a neural network, kinematics before integration, and a fine-tuned classifier before a 7B model is adapted or an agent is handed tools.

Medical-image segmentation on real data

U-Net segmentation on Kvasir-Instrument endoscopy images, evaluated against an untouched publisher test set. The score drop stays in the record; the recovery was earned without moving the threshold.

PyTorch NumPy Docker

Held-out Dice: 0.9226

Leakage-aware predictive maintenance

An evaluation-first pipeline on the public UCI AI4I dataset, using average precision for an imbalanced target and keeping preprocessing inside the split.

scikit-learn PyTorch pandas

Measured for imbalance, not accuracy theater

Explainable perception before deep learning

Classical CV detection, connected components, homography, and camera calibration built as a transparent baseline for later robotics work.

OpenCV NumPy geometry

Every transformation can be inspected

Vision-to-planning robotics integration

A simulation-first system connecting detections, transforms, safety gates, and planar planning—with safe rejections reported as outcomes, not hidden as errors.

vision planning simulation

The system is allowed to say “not safe”

The full sequence

Stage Project What it establishes
01 optimization-lab Linear and logistic regression from scratch in NumPy
02 medbot-vision Explainable detection, calibration, and geometry
03 surgiseg-ml U-Net segmentation on real endoscopy images
04 sensorguard-ml Leakage-aware evaluation on imbalanced public data
05 papertrail-rag Grounded retrieval over real ML and CS papers
06 textforge-finetune Fine-tuning against a meaningful baseline
07 loraforge-llm QLoRA adaptation with frozen evaluation
08 experiment-orchestrator A tool-using experiment agent with typed tools
09 wardnav A*, obstacle inflation, replanning, and noisy localization
10 surgiarm-sim FK/IK, collision checking, RRT, and pick-and-place
11 medbot-assist Integrated perception, safety, and planning
12 resume-radar Deterministic scoring and ranking automation

Full case studies, methodology, and evidence are on the portfolio.


Toolkit — Foundations before frameworks


Layer Tools
Languages Python · Java · JavaScript · HTML · CSS
ML & data PyTorch · scikit-learn · NumPy · pandas · OpenCV
Engineering Docker · GitHub Actions · pytest · Git · Linux
Working areas Computer vision · RAG · fine-tuning · robotics simulation · applied AI
Python, PyTorch, scikit-learn, OpenCV, Docker, Git, GitHub, and Linux

Activity / evidence of work

Mayank's GitHub profile activity Languages by repository Commits by time of day

Let's build something worth measuring — mghadia1@asu.edu


Open to summer, fall, and spring internships in medical AI, robotics, machine learning, and deep learning.

GitHub  ·  LinkedIn  ·  Email  ·  Portfolio


REAL DATA · HELD-OUT EVALUATION · HONEST WRITE-UPS · REPRODUCIBLE REPOS

Pinned Loading

  1. attentionforge attentionforge Public

    Attention from scratch, verified before it is optimized: NumPy and PyTorch references, a Triton online-softmax kernel with causal tile skipping and grouped-query attention, benchmarked honestly on …

    Python 1

  2. experiment-orchestrator experiment-orchestrator Public

    Tool-using ML experiment agent with seven typed tools and dataset protections enforced in code, not in the prompt

    Python 1

  3. kernelforge-cuda kernelforge-cuda Public

    Hand-written CUDA kernels for batched cosine-similarity + top-k retrieval, benchmarked honestly against cuBLAS and PyTorch

    Cuda 1

  4. loraforge-llm loraforge-llm Public

    QLoRA adaptation of Mistral-7B-Instruct on a single free T4, with calibration and one frozen held-out evaluation

    Python 1