Maclane Zich

Engineer building clinical AI. Penn Engineering.

Resume GitHub LinkedIn

2 apps on the App Store
Coca-Cola Scholar

Portrait of Maclane Zich in a navy suit and paisley tie on a white background

01 / Ventures

Ventures

Two iOS health apps, designed, built, and shipped as sole engineer. Both are live on the App Store.

Live on iOSAug 2025 to present

Nárm

Technical Co-Founder and Sole Engineer

Identifies fibers from a photo of a garment or its care label and flags skin irritants.

Workflow

01
Scan
Camera capture of a garment or its care label, in Label or Fabric mode.
02
Score
Fiber composition with scores for skin compatibility, climate, and sustainability.
03
Track
Scan history and a saved wardrobe with per-item scores.
04
Browse
Brand-agnostic tag reading and a reference for each fiber type.

Description

Users photograph a garment or its care label. Nárm returns the fabric composition, scored against the user's skin conditions, allergies, and climate.

A dual-path Gemini vision pipeline on Vertex AI separates label OCR from perceptual garment inference. Built in Flutter and Firebase. Patent pending; the basis of a Harvard Medical School capstone.

Flutter
Dart
Firebase
Gemini
Vertex AI
Computer vision
After Effects

label accuracy on held-out images: 98%
fiber-composition accuracy: 94%
downloads: 160+
launch-week conversion, 12 paid subscribers: 17.5%

Patent pending
Harvard Medical School capstone

App Store

Live on iOSFeb 2024 to present

HeadacheWise

Founder and Sole Engineer

Migraine patients log episodes and triggers; the app identifies which trigger combinations are associated with attacks.

Workflow

01
Log
Five-step entry: timing, pain, triggers, symptoms, and medications.
02
Review
Calendar view, color-coded by severity from mild to severe.
03
Learn
Top triggers by frequency and each medication's effect on episode duration.
04
Track
Episode totals, weekly summary, and recent episodes.

Description

Users log each migraine as an intensity time series rather than a single severity score and receive their own trigger patterns.

An on-device association-rule engine ranks trigger combinations by lift, identifying pairs associated with worse episodes together than either alone. A Bayesian time-series mode supports research analysis. The serverless Node.js backend handles batched push delivery, GDPR deletion cascades, and a scheduled job that closes abandoned episodes.

Flutter
Dart
Node.js
Cloud Functions
Association rules
Bayesian time series

users: 180+
episodes logged: 2,000+
Demoed at the U.S. Capitol, April 2026: Congressional App Challenge Winner

App Store headachewise.com

02 / Research & Experience

Research & Experience

Clinical AI research at Northwestern Medicine and Penn Medicine, and a software engineering internship at a psychiatric AI company.

01Northwestern Medicine

Institute for AI in Medicine

PaCo Research Platform

AI Engineer and Manuscript Co-Author

Feb 2026 to present

Voice AI that interviews patients about their medications and scores adherence across 4 clinical domains.

Contributed the LLM and analysis layers: transcripts mapped to the 4 QUILAM adherence domains with flags, supporting quotes, and confidence scores, behind a failover router across Gemini, OpenAI, and Llama.

inference cost cut: 60 to 80%
conversations / personas in the evaluation harness: 24 / 12

Manuscript in review
LLM evaluation
Voice AI
Gemini
OpenAI
Llama

02Perelman School of Medicine

Rajapakse Radiology Lab

Knee MRI landmark measurement

Undergraduate Researcher, Medical Imaging AI

Sep 2026 to present

The lab is developing a model that automates landmark measurement on knee MRI.

I annotate anatomical landmarks on knee MRI scans as one of the lab's annotators. The labeled scans become training data for the lab's model.

Medical imaging
Image annotation

03PsychNow

Psychiatric AI intake platform, Chicago

Patient-simulation suite

Software Engineer Intern

Mar to Aug 2025

TypeScript suite that tests a psychiatric AI's intake conversations end to end without live patients.

Three scripted personas (anxiety, bipolar disorder, postpartum depression) drive branching dialogue and simulated spoken input; a polling layer confirms each question exists in the backend before the test acts on it. Used to train psychiatrists on the AI intake and to demo the platform without patient data, including at Yale Medicine. Only high school engineer on the team.