HIGH-STAKES ML systems THAT CAN'T CHEAT

Aadi Sharma: SRM CSE undergrad (2024 to 2028, All India Rank 32, full merit scholar), research intern at CMU's XuLab. I work on RLVR (reinforcement learning with verifiable rewards), longitudinal 3D brain MRI across a 515+ patient ADNI cohort, the HEAT encryption-aware training framework, and concurrency-safe backends.

Aadi Sharma, ML systems and vision research intern at CMU XuLab
AADI · aadisharma.xyz CMU XULAB · ADNI
[ AIR 32 · FULL MERIT ] [ 515+ ADNI PATIENTS ] [ 32K CONTEXT · RLVR ] [ SRM CSE · 2024-2028 ]
Live System Switchboard TELEMETRY LIVE
Affiliation & Compute Telemetry
CMU XuLab · Prof. Min Xu
Research intern · RLVR + 3D vision
ACTIVE
Harvard Medical School
Taylor / Xiangrui Zeng · ADNI collaboration
ACTIVE
IIT BHU · Prof. Harsh Kashyap
HEAT framework · CKKS training
SHIPPED
UTSA · Hrishitwa Patel
Swarms + edge federated learning
ACTIVE
Training Stack
FSDP · ZeRO-2 · vLLM · DeepSpeed
ACTIVE
Selected Works // Quick Launch all repositories on GitHub
[R] LEDGERX

ACID Transaction Engine

Pessimistic row locking, idempotency, concurrent-transaction verification.

VIEW SOURCE
[R] CURRICULA

Curriculum Safety Verifier

4-level curriculum with self-correction trajectory augmentation.

VIEW SOURCE

What I work on

Four workstreams across three labs. The short version lives up front; the drawers carry what reviewers ask about.

TOKEN STREAM // AST GATE // QUARANTINE
Carnegie Mellon University

Continual RLVR Post-Training & Verifier Gates

Co-authoring "If-Taught Reasoners Forget How to be Safe" (under review): verifier gates and alignment replay buffers that cut Attack Success Rate across 10+ continual rounds, with Dr. Chhavi Yadav.

PyTorch FSDPDeepSpeed ZeRO-2vLLM
ACTIVE
CMU XuLab · Harvard Medical School

Longitudinal 3D Brain MRI & Atrophy Tracking

Built a cascaded 3D Siamese U-Net pipeline on a 515-patient ADNI cohort (70 GB) achieving 0.9971 intra-patient Elastic Dice with inter-patient deformation below 1e-4, with Taylor Zeng (CMU & Harvard Medical School).

PyTorch 3DEntropy-Gated CPS3D Siamese U-NetFreeSurfer
ACTIVE
IIT BHU · Varanasi · Nov 2025 – May 2026

HEAT: Training Under CKKS Encryption

With Prof. Harsh Kashyap: built a TenSEAL pipeline running neural inference directly on CKKS-encrypted data with secure aggregation, and developed HEAT (Token-Condensed Homomorphic Attention) for Vision Transformers.

TenSEALCKKS FHEVision TransformersSecure Aggregation
SHIPPED
UT San Antonio

Drone Swarms & Edge Federated Learning

First-author work across UTSA and University of the Arts London: 50-node edge swarms in DDIL networks serving Llama-3-8B, with 14.1x lower bandwidth than Epidemic routing and 99.6% decision preservation under 80% packet loss.

PyTorch GeometricGraph AttentionFederated LearningRaspberry Pi
ACTIVE

Papers, submissions, and work in progress

UNDER REVIEW
Task-Oriented Decentralized Semantic Synchronization for Swarm Resilience in Extreme DDIL Environments
First-author manuscript, under review. With Hrishitwa Patel (UTSA): straggler synchronization, non-IID local data, bandwidth-compressed updates.
SUBMITTED
Dynamic Spatiotemporal Graph Attention Networks for Functional Connectivity & Autism Subtyping
First-author manuscript, under review. Time-resolved connectivity graphs, attention over spatial and temporal neighbors.
UNDER REVIEW
HEAT: Token-Condensed Homomorphic Attention for Privacy-Preserving Vision Transformers
Under review (IIT BHU, 2026). With Prof. Harsh Kashyap.
CO-AUTHORED
Volumetric 3D ResNet-18 Classification of Autism Spectrum Disorder on ABIDE I fMRI
Co-authored study; the PyTorch pipeline is public for reproduction.
Repository
ACTIVE
If-Taught Reasoners Forget How to be Safe: Continual Post-Training Drift & RLVR Verification
With Dr. Chhavi Yadav (Carnegie Mellon University). Manuscript in progress.
ACTIVE
Longitudinal 3D Brain MRI Registration & Entropy-Gated Cross-Pseudo Supervision on ADNI
With Taylor Zeng (CMU & Harvard Medical School). 515+ patient ADNI cohort.

Things I've built

Three systems end to end, then the public bench. Repo descriptions are verbatim from GitHub.

W1QUEUED
W2QUEUED
W3QUEUED
W4QUEUED
W5QUEUED

LedgerX

High-concurrency ledger REST API that survives request spikes: pessimistic row locking, idempotency, and live concurrent-transaction verification.

PythonFastAPIPostgreSQLDocker

NeuraTwin

AI digital twin for longitudinal wellness: cognitive telemetry modeling and burnout detection, with the AURA conversational assistant.

TypeScriptNext.jsReactFastAPI
INGEST
TRAIN
EVALUATE
SERVE

SkillSync

Deployed ML platform for salary prediction, skill-gap analysis, and role demand forecasting, powered by Random Forest pipelines.

PythonScikit-LearnNext.jsVercel

Say hi

I'm in IST, my labs run on Eastern and Pacific time. The clock below is live, and everything copies with one click.

Languages

Python, TypeScript, JavaScript, SQL, HTML/CSS, Bash.

ML & Research

PyTorch, vLLM, DeepSpeed ZeRO-2, PyTorch FSDP, TensorFlow, TenSEAL, Scikit-Learn, OpenCV, MediaPipe, FreeSurfer.

Backend & Infra

FastAPI, Flask, Celery, Next.js, React, PostgreSQL, Docker, Linux, Git, Vercel.

EMAIL aadisharma2808@gmail.com
LINKEDIN linkedin.com/in/aadisharma11
GITHUB github.com/AadiSharma1
SITE aadisharma.xyz CV ↓
IST 00:00:00 ET 00:00:00 PT 00:00:00

Message composer

Pick a scope, scribble the problem, and this opens a pre-filled draft in your mail client.