ShauryadityaSingh

Applied AI Engineer

I build trustworthy AI for climate, energy, and safety-critical systems. My work centres on uncertainty, abstention and selective prediction. I also build film-AI tools for roto, cleanup and restoration.

View projectsRead the research

Open to full-time roles from January 2027; graduating May 2027.

About

I'm an applied AI engineer and researcher. I'm finishing an M.S. in Applied Artificial Intelligence (Data Engineering concentration) at Stevens Institute of Technology, on the thesis track with Prof. Lei Wu.

My research is trustworthy AI for climate, energy, and safety-critical systems: uncertainty, abstention and selective prediction. I also build film-AI tools for roto, cleanup and restoration.

My background is process engineering (B.A.Sc., University of British Columbia). I have shipped data and ML work for a national supportive-housing nonprofit, a textile exporter and a fintech.

Open to full-time roles from January 2027; graduating May 2027.

Portrait of Shauryaditya Singh
Education
M.S. Applied AI, Stevens Institute of Technology
Aug 2025 to May 2027 (expected)
B.A.Sc. Process Engineering, University of British Columbia
Completed 2025
Research
2 published papers, 1 under review
Fellowship
Provost Master's Fellowship ($15,000)
Based in
Jersey City, NJ (NYC metro)
Languages
English (professional), Hindi (native)
Published papers
2
incl. 1 Best Paper Award (ICICC 2026)
Regions × years of ERA5 hours
12 × 47
benchmarked in HeatBench, 1979 to 2025
AUROC, hallucination detection
0.85
with a 95% precision guarantee (Selective RAG)
On-device dictation latency
283 ms
Hush, p50

Projects

Five in depth, each with a demo you can use.

  1. Building

    CleanPlate + RotoBench

    An open-source, fully local roto and cleanup tool for film and VFX, plus RotoBench, an independent benchmark of whether AI mattes are production-ready.

    hair-region matte error with the HQ hair mode
    -37.6%
    earlier claims hold on RotoBench re-test
    8 of 13
  2. In preparation

    HeatBench

    A benchmark of selective prediction and conformal uncertainty for heatwave early warning under distribution shift.

    coverage gap against score-KS, a straight line
    R² 0.919
    region pairs break the bound |gap| ≤ KS
    0 of 132
  3. In progress, phase 2

    AnswerPrint + AnswerPrint Bench

    An open-source restoration workbench for classic film, plus a public benchmark of which restorations humans actually accept.

    rights-cleared clips from 28 sources
    92
    sources refused at screening
    9 of 37

  4. Concept simulation, synthetic day

    In progress

    Data-centre microgrid scheduler (M.S. thesis)

    Real-time scheduling of AI compute jobs on a data-centre campus microgrid under solar and electricity-price uncertainty.

    between re-plans
    5 min
    of lookahead
    2 to 4 h
  5. Done

    Selective RAG with Conformal Abstention

    A hallucination detector for RAG that abstains when it is unsure.

    AUROC on RAGTruth, +0.24 over NLI
    0.85
    precision held at the 95% target
    95.2%

Smaller builds

  • In daily use

    Hush

    On-device dictation for macOS that replaced Wispr Flow.

    end-to-end latency at p50
    283 ms
    whispered-speech WER, from 9.47%
    5.35%

    Private repo

    • White (reference)1.00, meets the four-fifths rule
    • Asian0.97, meets the four-fifths rule
    • Black0.78, below the four-fifths line
    • Native American0.73, below the four-fifths line
    • Pacific Islander0.71, below the four-fifths line

    Approval rate against White applicants. The line is the four-fifths rule (0.8); bars below it fail.

    Done

    FairLens: HMDA fairness audit

    A forensic audit of algorithmic bias in US mortgage lending.

    A paper on this audit is in revision.

    HMDA records audited
    801,595
    accuracy while meeting the four-fifths rule
    94 to 96%

    Data: the 2024 HMDA Public Loan/Application Register, published by the FFIEC.

    Live app: FairLens: HMDA fairness audit
  • Done

    AutoTabber

    Audio to guitar tab via pitch detection and onset analysis.

    Co-built with Daniel Ahn.

    Try the demo: AutoTabber

Archive

Older and paused work.

Research

Trustworthy AI for climate, energy, and safety-critical systems.

  1. ICICC 2026 Springer

    Published

    Best Paper Award

    Reinforcement Learning-Based Energy Dispatch for AI-Enabled Surgical Units under Variable Renewable Supply

    An RL agent learns to dispatch renewable energy to surgical ICUs, balancing supply variability with hard uptime requirements on life-critical equipment.

    Random Forest, XGBoost, LSTM, GRU and 3 more

    Read on SpringerLink

  2. AIMLA 2026 IEEE

    Published

    Digital Twin for Battery-Powered Medical IoT Fleet: Safety-Aware Energy Budgeting Across Sensors, Edge AI, and Data Transmission

    A Digital Twin layer allocates battery power across sensing, edge inference, and transmission via a Clinical Urgency Score, with physics + ML predicting battery state of health.

    Digital Twin Modeling, Safety-Aware Energy Budgeting (SAEB), Physics-based Battery Modeling, ML-based SoH and RUL Prediction and 3 more

    Read on IEEE Xplore

  3. AIES 2026 / ICMLT 2026

    Under review

    When Should AI Abstain? A Decision-Theoretic Framework for Uncertainty-Aware Human Activity Recognition in Safety-Critical Settings

    Sole-authored research on calibrated abstention in human activity recognition for high-stakes deployments.

    Decision Theory, Uncertainty Quantification, HAR, Calibration and 1 more

  4. arXiv (planned)

    In preparation

    HeatBench: selective prediction and conformal uncertainty for heatwave early warning under distribution shift

    A benchmark of selective prediction and conformal uncertainty for heatwave early warning under distribution shift, across 12 regions and 47 years of ERA5 data.

    Split Conformal Prediction, Selective Prediction, Distribution Shift (Wasserstein, KS), ERA5 Reanalysis and 2 more

Experience

Applied data and ML across housing policy, operations, and fintech.

  1. 2026Data Science InternCorporation for Supportive Housing (CSH)May 19 to Aug 27, 2026Built the state-by-state LIHTC Qualified Allocation Plan (QAP) extraction pipeline that codes every state's affordable-housing scoring rules into one analysis workbook: 57 of 57 jurisdictions covered in the final build, with per-cell confidence shading and a documented-sources rule. Red (divergent) cells fell from 209 to 25 across builds v4 to v7. Extraction is OCR plus deterministic parsing.
  2. 2025Data Science InternUnified MentorMar 2025 to Aug 2025Delivered four end-to-end ML projects across HR, mobility and supply chain, each deployed as a Streamlit app for non-technical users.
  3. 2023 to 2025Data AnalystSINI Designs Pvt. Ltd.Jan 2023 to Jan 2025Built Tableau production and supplier KPI dashboards used daily for operations monitoring.
  4. 2022Business Development AnalystNeo FinancialMay 2022 to Aug 2022Market and sales-performance analysis that identified growth opportunities; contributed to a 5% increase in sales volume.

Skills

  • Python
  • SQL
  • TypeScript
  • Swift
  • C++
  • PyTorch
  • scikit-learn
  • XGBoost
  • Hugging Face Transformers
  • Conformal and selective prediction
  • Uncertainty quantification
  • Reinforcement learning
  • Time-series forecasting
  • Computer vision
  • pandas
  • NumPy
  • xarray / ERA5 pipelines
  • DuckDB
  • Postgres (Neon)
  • SQLite
  • Excel
  • Tableau
  • Power BI
  • Next.js
  • React
  • Expo / React Native
  • Flask
  • Gradio
  • Streamlit
  • Vercel
  • Azure App Service
  • Modal
  • Kaggle
  • Hugging Face Hub
  • FFmpeg

If something here resonated, let’s talk.

shauryaditiyas@gmail.com

Open to full-time roles from January 2027; graduating May 2027. Also open to research collaborations.