Aafi Mansuri

    _        _     _____  ___ 
   / \      / \   |  ___||_ _|
  / _ \    / _ \  | |_    | | 
 / ___ \  / ___ \ |  _|   | | 
/_/   \_\/_/   \_\|_|    |___|
 __  __     _     _   _  ____   _   _  ____   ___ 
|  \/  |   / \   | \ | |/ ___| | | | ||  _ \ |_ _|
| |\/| |  / _ \  |  \| |\___ \ | | | || |_) | | | 
| |  | | / ___ \ | |\  | ___) || |_| ||  _ <  | | 
|_|  |_|/_/   \_\|_| \_||____/  \___/ |_| \_\|___|

I move between research, modeling, evaluation, and engineering.

Right now, I’m doing research on a cardiovascular-risk model at Emory’s HITI Lab.

In a rush? prints the whole record on one screen: roles, projects, education, contact.

All of that is the work. There is also a record wall, a shelf of paperbacks and a CRT full of games, filed under .

experience.log3 rolesEmory, BISAG-N, iBhavan
education.log2 degreesNortheastern, Ganpat University
projects/4 projectsmultimodal fusion, T5 fine-tune, OMR pipeline, text-to-SQL
skills.json6 groupsML, statistics, medical imaging, full-stack
offline5 thingsmusic, reading, games, photography
contact4 waysemail, phone, LinkedIn, GitHub

Click anything above, or type it below. Tab completes, help lists everything.

Experience

4c1f9abHEAD -> main
Data Scientist
Healthcare Innovations and Translational Informatics Lab, Emory University May 2026 to present · Atlanta, GA
Evaluating test-retest reliability and longitudinal stability of a deep learning model that quantifies breast arterial calcification (BAC), a cardiovascular risk biomarker, across four mammography modalities using Emory’s Breast Imaging Dataset (EMBED).
- BACcardiovascular risk biomarker
- statisticsEngineered a statistical evaluation framework, combining Bland-Altman analysis, repeatability coefficients, Cohen’s kappa, and confusion-matrix agreement metrics, to isolate measurement noise from true physiological change.
- survival analysisDesigning a Cox proportional-hazards survival analysis testing whether longitudinal BAC progression improves prediction of major adverse cardiac events (MACE) beyond a single baseline measurement, built on a landmark-anchored cohort of 25,794 women with 1,228 observed events.
- engineeringRan deep learning model inference and processed large-scale DICOM and mammography metadata in Python, supporting statistical analysis and radiologic review via MD.ai.
b82a417
AI Engineer
Bhaskaracharya Institute for Space Applications and Geo-Informatics Jan 2025 to Apr 2025 · Gandhinagar, India
Engineered an AI-powered web app enabling natural-language database queries via a self-hosted Gemma-2 27B LLM through Ollama.
- 27Bparameter LLM, self-hosted
- full-stackBuilt a full-stack system (Next.js, FastAPI, PostgreSQL) with role-based access control (Clerk) and an administrative interface for multi-database schema management.
- modelingImproved text-to-SQL accuracy through schema-grounded prompt engineering and dynamic schema mapping across heterogeneous database schemas.
1d5e6f0
Software Developer
iBhavan May 2023 to Jun 2023
Built a cross-platform mobile application using Flutter and Dart, connecting clients with in-home nurses for healthcare services, including appointment booking and real-time data synchronization.
- engineeringIntegrated Firebase Authentication and Cloud Firestore into the Flutter app for secure login and structured data storage.
- interfaceDesigned and built responsive Flutter UI components for a consistent cross-platform user experience.

Education

e07d33cHEAD -> ms-cs
MS, Computer Science
Northeastern University Sep 2025 to Dec 2027 · Boston, MA
- 4.0GPA
coursework
- Natural Language Processing
- Programming Design Paradigms
- Algorithms
- Pattern Recognition & Computer Vision
- Foundations of Generative AI
9a30c74
B.Tech, Computer Science and Engineering
Ganpat University Aug 2021 to May 2025 · Mehsana, India
- 3.85GPA
coursework
- Machine Learning
- Deep Learning
- Artificial Intelligence
- Predictive Modeling
- Cloud Computing
- Functional Programming
- Data Structures
- DBMS

Projects

total 4

hiti-datathon-2026-cxr-mammo/

Mammography + Chest X-Ray Risk Subtyping

HITI Lab Datathon · August 2026

Multimodal fusion beat the best single-modality baseline on AUROC for three of the four outcomes, identifying prevalent cardiometabolic disease from screening images taken for something else entirely.

Developed a multimodal fusion pipeline combining anonymized mammography and chest X-ray foundation-model embeddings to identify prevalent type 2 diabetes, hypertension, hyperlipidemia, and chronic kidney disease from routine screening imaging.
Evaluated L2-penalized logistic regression against early-fusion, late-fusion, and gated-fusion MLP architectures, using RankMe-based linear probing for embedding screening and auditing performance across age, race, BMI, and imaging interval.
Demonstrated the potential for opportunistic cardiometabolic screening from imaging exams already performed for other clinical purposes, with multimodal fusion improving identification of type 2 diabetes and chronic kidney disease without requiring additional imaging.

repository

                ......:::::;;;;;ooOOOooOOOOoooo;;;::::........   ..           
           .......:::::::;;;;;ooooOOOOooOOoooooo;;;::::::::......:::          
       .....:::::::::::;;;ooooooooOOooooOOoooooooo;;;::::::::::::;;;....      
 .....:::::::::::::::;;;;;;ooooooOOOoo;oOooooo;;;;;;;;;::::::::::ooo:::::.... 
::::;;;;;::::;;:::;;;;;:;;;::::;oooOooooOOoo;;::::::::;;;;;;;;;;;;;;;;;;;;::::
;;;;;;oooooooo;;;oo;:::::::::::;;oOOooooOoo;;::::::;:;;oooooooooooooooo;;;ooOo
Ooooooo;;oooOOoOOooo::::::...::;;;oOoooOOoo;:::::;;;;;;;oooooo;;oOoooooOOOOOOo
OOOooooooOooooooooo;;;::;;:::::::;oOoOoOOo;;:;;;;:;;::::::;;oooooooOOOOOOOOOOo
OOOOOOO8OoooooOo::::::::::::;;;;;;oOoooOOOo;;;:::::..:::::::;ooooooOO8OOOO8OOo
888O88O8OooooOo:::::::...::::::;;oooOOOOOOo;;:::::::::::::::;;OOooOOO8O8888OOO
O888OOOOOoooOo;;::::::::.:::::::;;oOooOOOo;:::;:::::.::..:::;;oOOoOOOOOO888OOO
888OOOOOOooOOo;;:::........::::;;;;oooOOOo;;;;;:..........:::;oOOOOOOOOOO888OO
888OOOOOOoOOoo;:::...........::::;ooOOOOOOo;:::.........:::::;;oOOOOOOOOO8888O
888OOOOOOOOOo;::::::.........:.:::;;oOOOOOo;;::...:....:::::;;;;OOOOOOOOO88O8O
O88OOOOOOOOO;;;:::::..:...::::::::;ooOOOOOooo;:::::......:::;;;oO8OOOO88O888OO
888OOOOOOOOOo;;:::::........:::::;;oOOOOOOOoo;:::........:::::;oO8OOOO888888OO
888OO8OOOOOOo;:::::..........::::;;oOOOOOOoo;;:.::..::::::::::;;oO88OOO88888OO
8888OOOOO88o;;::::::::.::.:::::;;;;oOOOOOOooo;;::::::::::::::;;ooO88OOOO8888OO
8888OOOOO8Oo;;;;::::::::::::;;:;;;ooOOOOOOOooo;;;::....::::::;;;oO888OOO8888OO
888OOOOO88Oo;;;;:::.......:::::;;;oOOO8OOOOOoo;;;;:::..::::::::;oOO88OOOO88OOO
O88OOOOO88Oo;;:::::........:::::;oOOOO8OOOOOoo;;;;;::::::::::::;;oO888OOO8OOOO
OO8OOOO888Oo;::::::::.::::::;;;;;ooOOO8888OOOo;;;;;;::;::::::;;;;oo888OOOOOOOO
OOOOOO8888o;;;::::::::::::;;;;;;;ooOOO88888OOO;:::::::::::::;;;ooooO88OOOOOOoO
OOOOOO888Oo;;;;::::::::::::;;;;;;ooOO8888888OOo;::;::::::::::;;;ooOO88OOOOoooO
OOOOOO888Ooo;;;::::::::::;;;;:;;;oO8888888888OOoo;;;;::::::::;;;;oO8888OOooooo
OOOOO8888OOo;;:::::::::::;;;;;;;oOO88888888888OOoo;;;;:;;;;;;;:;;;oO888OOooooo
oooOO8888Oo;;::::::::::::;;;;;;ooOO88888888888OOOoo;:::::;;;;;;;;;oO88OOOoooo;
oooOO8888oo;;::::;:;:::::;;;;;;oOO8888888888888OOOOo;:;;;;:::;;;;ooo88OOOooo;;
ooOOO888Oo;;;;;:::::::;;;;;::;;oOO8888@88@88888888OOo;::;:::::;;;ooOO8OOOoo;;;
oooOO888Ooo;;;;::::::::::;;;;;OOO88888@@@@@@@888888OOo::::;::;;;;;;oO88OOoo;;;
oooO8888Oo;;;::;::::::::;;;;;;oOO8888@@@@@@@@8888O88Oo;;:::;;;;;;;;oO88OOooo;:
;ooO8888Oo;;::::;:;::::;:;;;;;oO888@@@@@@@@@@888888OOOo;;;;;;;;;;;;;o88OOoo;;:
;ooOO88Oo;;;;::;:;oooooooo;;ooOO888@@@@@@@@@@@888OOOOoooo;;;;;;:;;;;;O8OOoo;;:
;ooOO88o;;;;;;oOOOOOOOOOOOO88888888@@@@@@@@@@8888OOOOoOOOOOooo;;;;;;;o8OOo;;;:
;ooOO8Oo;:;oOO8O8888888888888888888@@@@@@@@@@88888888OOOOOOOOOOOOoo;;oOOOo;;::
;;ooOOo;;OOO8888888888888888888888@@@@@@@@@@@888888888888888888OOOOOoooOoo;;:.
:;;oOOoo88888888888888@@@@@@@@@8@@@@@@@@@@@@@@@@@88888@@88888888OOOO8OoOOo;;:.
:;;oOOO8888888@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@888888888O8OOOOOo;::.
.:;oO8O88888@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@888@@88888OOOOoo;::.

PaperChord/

PaperChord

AI chord progression generator

Fine-tuned a T5-small model on a genre-conditioned synthetic dataset encoding music-theory rules to generate Roman-numeral chord progressions from natural-language mood prompts.
Built a full-stack app (Next.js frontend, FastAPI backend) with in-browser MIDI preview (Tone.js) and multi-style, downloadable MIDI export.
Designed a rule-based music generation pipeline that combines neural generation with music-theory constraints, including key transposition, voice leading, and stylistic variation to produce musically coherent progressions.

repository

SheetVision/

SheetVision

Optical music recognition pipeline

End-to-end OMR pipeline using OpenCV for staff line removal and symbol segmentation, with pitch mapping from vertical positions to treble clef indices.
Trained a 7-class PyTorch CNN on a synthetic dataset of 896 SMuFL-rendered symbols with rotation, translation, and noise augmentation.
Engineered the complete pipeline from synthetic data generation and symbol classification to pitch reconstruction and MIDI output, enabling recognition of real printed melodies without relying on manually labeled training data.

repository

AI-Text-to-Sql/

AI-Driven Text-to-SQL Query Assistant

Natural-language querying over live databases

Full-stack assistant that turns a plain-English question into SQL, runs it against a dynamically selected PostgreSQL database, and returns the result set in a chat interface.
Inference runs entirely locally on a Gemma-2 model served through Ollama and specialised with a custom Modelfile, grounded on a combined-view DDL rather than raw table definitions, so generated SQL stays inside the schema the caller is allowed to see and no query ever leaves the deployment.
Next.js frontend and FastAPI backend, with Clerk for role-based access and an admin panel for registering databases and views.

repository

Skills

Machine learning & AI

core

PyTorch
Hugging Face Transformers
Scikit-learn

also

OpenCV, TensorFlow, YOLO, local LLMs (Ollama)

used in

T5-small fine-tune for PaperChord, 7-class CNN for SheetVision
Self-hosted Gemma-2 27B for text-to-SQL at BISAG-N

Statistics & evaluation

core

Bland-Altman analysis
Cohen’s kappa
repeatability coefficients

also

cohort design, Pandas, NumPy, Tableau

used in

Evaluation framework isolating measurement noise from physiological change, HITI Lab
Cox proportional-hazards survival analysis, 25,794-woman EMBED cohort

Medical imaging & health data

core

DICOM processing
model inference pipelines

also

mammography metadata (EMBED), MD.ai

used in

Large-scale DICOM and mammography metadata in Python, HITI Lab
Mammography and chest X-ray embedding fusion, datathon

Web & full-stack

core

Next.js
FastAPI
React.js

also

Tailwind CSS, Angular, HTML5 / CSS3

used in

Next.js, FastAPI and PostgreSQL app with role-based access control, BISAG-N
PaperChord frontend, API and MIDI export

Languages

core

Python
JavaScript / TypeScript

also

SQL, Java, C++

used in

Python across research, modeling and evaluation; TypeScript across everything shipped to a browser

Data stores & tooling

core

PostgreSQL
Git

also

MongoDB, Firebase, AWS, JupyterLab

used in

Dynamic schema mapping across heterogeneous PostgreSQL databases, BISAG-N
Firebase Authentication and Cloud Firestore, iBhavan

Offline

The things I do when nothing is being evaluated.

There is a lot of overlap. PaperChord and SheetVision both exist because I make music and got curious whether a model could hear what I hear.

Making and listening to music, playing games, reading fiction, and taking photographs.

Making music

Producing is the hobby I keep coming back to. Mostly house, the groovy kind, wandering into pop when a melody asks for it and usually landing somewhere dreamy or a bit dark. All of it built in FL Studio.

Releases on Spotify · Remixes and reels

Listening

Constantly, and across more genres than is probably defensible.

on repeat

DÁKITIBad Bunny and Jhay Cortez
Tell Me The TruthTwo Feet
Carry YouMartin Garrix & Third ≡ Party
Pull Me Through The Fire (Citadelle Remix)Matisse & Sadko

Reading fiction

Crime and legal thrillers first, science fiction when I want the bigger question, and lately I have been working my way through LitRPG, which I did not expect to like as much as I do.

shelf

The ClientJohn Grisham
A Time for MercyJohn Grisham
NyxiaScott Reintgen
Nyxia UnleashedScott Reintgen

Games

Single-player, mostly. Something with immersion, a story worth staying for, and enough depth that the systems keep opening up. Quitting competitive grinding turned out to be the closest thing to inner peace I have found.

recently

Clair Obscur: Expedition 33
Red Dead Redemption II
The Last of Us Part I
Kingdom Come: Deliverance

Photography

Just moments, mostly on my phone. I like chasing a cinematic mood out of whatever light is already there, and nudging it until it looks the way the moment felt.

Contact

Before you go, there is one more command worth running:

Aafi Mansuri

Contents

Experience

Data Scientist

AI Engineer

Software Developer

Education

MS, Computer Science

B.Tech, Computer Science and Engineering

Projects

Mammography + Chest X-Ray Risk Subtyping

PaperChord

SheetVision

AI-Driven Text-to-SQL Query Assistant

Skills

Machine learning & AI

Statistics & evaluation

Medical imaging & health data

Web & full-stack

Languages

Data stores & tooling

Offline

Making music

Listening

Reading fiction

Games

Photography

Contact