Shaurya Attreya

2026—present

CrackTwin-Trust

Can a crack-tip estimator recognize when it should not be trusted?

Engineers use camera measurements to follow fatigue cracks in metal. A model can automate that work, but a confident, wrong prediction is dangerous. I’m studying how image models, fracture-mechanics fits, and an independent crack-length sensor can check one another—and when the system should stop and ask for review.

0.801 mm: median error
333 / 334: predictions within 2 mm
167: external acquisitions

The current 64-pixel model reached these results without training on the external DLR experiment. A disagreement check between the image model and the physics estimate also identified the one prediction outside the 2 mm target. These are development results, not a final claim: the last benchmark split remains sealed while I freeze the method and prepare a second independent test.

How the project changed

I began by trying to make a physics-informed neural network recover crack position directly. It performed poorly: the inverse problem did not contain enough information in each small measurement window. That failure changed the research question. Instead of forcing one model to solve every case, I now compare several estimators, measure what information is actually present, and reject unsafe predictions.

A second surprise came from damaged inputs. On clean benchmark data, the image model was almost perfect. Under structured occlusion it could fail confidently, while simple neural confidence and scalar physics residuals were both close to chance. The useful signal came from disagreement with an independent sensor—not from assuming that any single model knew its own limits.

Python · PyTorch · scientific computing · fracture mechanics · DIC

Selected work

Things I’ve built

Internship · Frontend & AI July 2026—present

OpenGraph AI

Frontend designer and AI engineer at OpenGraph AI, an open-source platform that turns tables, text, images, audio, and video into queryable knowledge graphs for AI agent reasoning.

I designed and built the full public-facing landing page end to end—marketing narrative, an interactive graph-exploration playground, and a guided pipeline walkthrough—on React 19, TanStack Router/Start/Query, Tailwind, and shadcn/ui, with Supabase-backed auth and graph export. It isn't published yet; I'm opening my first PRs next. I've also contributed to the project's Model Context Protocol (MCP) tooling for opengraph-image, which exposes five MCP tools so any MCP-compatible agent can build and reason over an image-derived knowledge graph conversationally.

React 19 · TypeScript · TanStack · Tailwind · shadcn/ui · Supabase · Python · MCP

Hardware · Air quality vivo Ignite 2026 · Achiever 30, Ed. 4

Aether Smart Badge

A wearable air-quality monitor for outdoor workers. The badge measures the PM2.5 a person actually encounters, keeps track of cumulative exposure, and gives a physical warning through a haptic motor. A phone connects over Bluetooth and adds a server-generated 72-hour pollution forecast.

₹4,433: core badge parts
142 / 142: app tests passing
51.14: +24 h MAE (µg/m³)

The first badge is assembled. An ESP32-C3, an SPS30 particle sensor and a BME280 sit in a 3D-printed enclosure, and the badge's own display shows the air-quality band, live PM2.5, running shift dose, temperature and humidity, with no phone required. The companion app is an installable PWA, also packaged for Android, and the FastAPI server that feeds it passes its 15 API tests.

Prototype Glimpse Select to inspect full size

Hardware Assembled badge with live reading and shift dose

App Glimpse Select a screen to enlarge

Aether app pairing screen: Meet Aether, Air you can act on, a disconnected badge, a Pair badge button, and a Private by design note saying there is no account and no cloud history

Pair Private by design

Aether app Now screen in demo mode with sample data: PM2.5 of 78 µg/m³ in the Moderate band, advice to limit exposure, shift dose at 64% of a 192 µg·h/m³ reference budget, and 138 µg/m³ expected in the next two hours

Now Live reading & shift dose

Aether app Forecast screen in demo mode with sample forecast data: mask-hour tiles for now, +1 h and +6 h, and a 72-hour outlook chart for Delhi-NCR with a shaded error band

Forecast Mask hours & 72-hour outlook

I froze the evaluation rules before training. The deployable gradient-boosted model clears the +24-hour gate on an untouched test winter (51.14 MAE and 84.62 µg/m³ RMSE, versus persistence at 51.86 and 99.46) and remains the forecast used by the demo.

The retrained residual Fourier Neural Operator is a partial success: on the anchored evaluation it beats persistence and climatology at +6, +48, and +72 hours, but not at +24 hours. I report the winning and losing horizons explicitly; the next model change is a trained origin-observation channel, not an unsupported production claim.

Still open: the sensor has not been calibrated against a reference monitor, so dose is a relative measure for now; the forecast in the demo is a labelled historical hindcast, not live data; and no badge has yet run a full shift on a worker. Calibration and a small field pilot come next.

ESP32-C3 · C++ · BLE · Python · FastAPI · PWA · CadQuery · 3D printing · forecasting

Medical imaging · Competition In progress

RSNA Knee MRI Abnormality Detection

A system for predicting twelve knee abnormalities from multi-series MRI scans. The dataset contains 4,407 studies but only 58 with expert labels, so I used multilingual radiology reports to create soft supervision for the remaining scans, then trained a 2.5D vision transformer across sagittal, coronal, and axial views.

0.7835: pooled OOF macro AUC
+0.0973: AUC over the first version
819,078: training DICOMs audited

With the imaging architecture held fixed, a stronger three-source report-label ensemble improved every validation fold. Its label AUC was 0.8956 on the common 57-study expert subset, while five-fold mean imaging AUC rose from 0.7239 to 0.8125. The 0.7835 figure is the strict pooled out-of-fold result across all 58 expert-labeled studies. A competition leaderboard result is still pending, and this is a research system—not a clinical diagnostic tool.

Notebook & Model Glimpse Select to inspect full size

Jupyter notebook output of 8 knee MRI sagittal slices aspect-fitted

DICOM Processing Cached series slices (aspect-fitted)

Jupyter notebook plots of loss convergence, binary accuracy, and macro AUC progression

Gate Verification Loss, accuracy & macro AUC gate

Jupyter notebook cell inspecting tensor shapes and setting overfit gate requirements

Pipeline Setup Study cache inspection & gate criteria

Jupyter notebook training logs from epoch 1 to epoch 160 showing convergence

Training Logs 160-epoch memorization progress

Plots of gold fine-tuning loss and held-out gold macro AUC across epochs

Model Evaluation Gold fine-tuning loss & held-out macro AUC

PyTorch · DINOv2 · DICOM · weak supervision · five-fold validation

Web · Physics Complete

Physiverse

A browser-based physics lab for learning by changing a system and watching it respond. I built interactive simulations that connect equations with motion, parameter sweeps, and visual experiments.

40: interactive simulations
8: subject areas
3: numerical solvers

The library spans mechanics, fluids, electromagnetism, optics, thermodynamics, waves, quantum mechanics, and introductory machine learning. Simulations include live measurements, graphs, guided experiments, and data export.

Simulation Glimpse Select to inspect full size

Interactive 3D crank-slider four-stroke engine simulation with live PV diagram

Thermodynamics & Mechanics 4-stroke engine & live PV indicator

Quantum wave packet tunneling with 3D waterfall history and transmission curves

Quantum Mechanics Wave packet barrier tunneling & waterfall

Fluid dynamics simulation showing Kármán vortex street velocity field

Fluid Dynamics Kármán vortex street velocity field

Physiverse Lab interactive simulation library across 8 subject areas

Library 40 interactive modules across 8 domains

JavaScript · Canvas/WebGL · RK4 · adaptive RK45 · Yoshida integration

Education · AI Prototype

Open-response marking engine

A tool for comparing written student answers with a teacher’s rubric. I combined language-model feedback with semantic matching and fixed validation rules to make the scoring more consistent and easier to audit.

Python · language models · REST APIs

Outside the work

Other things I enjoy

Soccer

I play soccer—a welcome change of pace from long hours spent building and debugging.

Chess

I enjoy chess and the quiet work of thinking through a position before making a move. I represented Kurukshetra at the zonal/district and state level in the 2024–25 season.

On my shelf

These books have helped shape how I think, work, and approach difficult problems.

Can’t Hurt Me David Goggins
Artificial Intelligence: A Modern Approach 4th edition · Stuart Russell and Peter Norvig

Away from the screen

Learning with others

School STEM Hub

I founded a school group for hands-on physics, introductory machine learning, and Arduino projects. I write workshop material and help younger students turn classroom ideas into working prototypes.

Scientific computing community

I also coordinate an online student community for physics simulation and machine learning, with peer reviews, shared technical resources, and small open-source collaborations.

Background

A little more

I study at M. S. Senior Secondary School in Jhansa. I work mostly in Python, Julia, JavaScript, and C/C++, with tools including PyTorch, JAX, WebGL, and Arduino. My interests sit between computational mechanics, scientific machine learning, and building better ways to explore difficult ideas.