Devansh Kumar

+91 7007484933 · work.devanshkumar@gmail.com · LinkedIn · GitHub · LeetCode · Google Scholar · ORCID · devansh.qzz.io
Published at IEEE InterSys 2026 · First-author manuscript under review (IEEE Access) · 284K+ transactions · ROC-AUC 0.9851
Summary
B.Tech CSE student at Bennett University (CGPA 8.75) with hands-on experience in backend engineering, asynchronous task pipelines, and applied machine learning. Co-author of published IEEE conference paper on computer-vision pose estimation for gait analysis (InterSys 2026); first-author of fraud decisioning paper (284K+ transactions, ROC-AUC 0.9851) under review at IEEE Access. Developer of production-grade compliance scanning and cryptographic pipelines using FastAPI, PostgreSQL, and Redis. Co-authored published book; IEEE BU Chairperson.
Education
Bennett University, Greater Noida Aug 2024 – Aug 2028
B.Tech – Computer Science & Engineering  |  CGPA: 8.75 / 10.0
Coursework: DSA, OOP, DBMS, OS, Computer Networks, Software Engineering
Sunbeam School Varuna, Varanasi 2016 – 2023
Grade 10: 95%  |  Grade 12: 89% — Physics, Chemistry & Mathematics
Experience & Research
Undergraduate Research  ·  Bennett University 2025 – 2026
  • Co-authored methodological review on camera-based pose estimation for neurodegenerative gait monitoring (IEEE InterSys 2026); led comparative analysis and literature synthesis, structuring vision-based kinematic methods into an evaluation framework.
Software Developer Intern  ·  BuildIt Service Mar 2026 – Present
Gurugram, Haryana
  • Built lead generation automation pipelines and REST API integrations processing 500+ leads/week; developed internal dashboards with SQL-driven analytics using Python; shipped client-facing web apps across agile sprint cycles.
Application Developer (Flutter)  ·  Raahi Oct 2025 – Mar 2026
Noida, Uttar Pradesh
  • Worked with an engineering team to build and launch a cross-platform mobile app (iOS, Android, Web) using Flutter and BLoC; implemented secure authentication and SQLite local persistence across 4+ agile sprint releases.
Chairperson  ·  IEEE Student Branch, Bennett University Sep 2025 – Present
  • Lead 100+ member branch; secured ₹1.48L+ in sponsorships; organised Global AI Summit 2.0 (3,000–4,000+ attendees) and Triverse 3.0 (600+ participants). Led branch to 3rd place — IEEE UP Section Best Emerging Student Branch Award 2025.
Projects
SatyaLabel — Legal Metrology Label Compliance Scanner GitHub
Python · FastAPI · Tesseract + EasyOCR · OpenCV · Celery + Redis · PostgreSQL 16 + PostGIS · SQLAlchemy 2.0 async · ReportLab · Docker · 102 tests
  • Built a camera-scan compliance checker for packaged-commodity labels under the Legal Metrology (Packaged Commodities) Rules, 2011: dual-engine OCR (Tesseract primary, confidence-triggered EasyOCR fallback below 0.60), OpenCV pre-processing (deskew, de-glare, perspective correction), and extraction of 10 mandatory declarations with per-field confidence scores.
  • Engineered an explainable rule engine where every verdict cites the specific violated rule (e.g. Rule 6(1)(f)), with cross-field validation (e.g. Best-Before must postdate Manufacture) and court-ready PDF evidence reports embedding scan photos via ReportLab.
  • Shipped async scan pipeline (Celery + Redis, 202-accepted + polling), geo-tagged scan persistence on PostGIS with GiST-indexed violation heatmaps, JWT auth with citizen/inspector/admin roles, and CI-backed Docker Compose deployment.
CryptoFlow — Multimodal Medical Data Encryption & Cross-Modal Integrity Binding github.com/devantaris/cryptoflow
Python · AES-256-GCM · HMAC-SHA-256 · RSA-OAEP · Dempster-Shafer · Deep Evidential Learning · Typer CLI · pytest (76)
  • Engineered a 6-stage cryptographic pipeline for multimodal patient bundles (DICOM, radiology reports, EHR metadata) achieving 120+ MB/s encryption, 130+ MB/s decryption at <0.1% storage overhead; custom binary container format with cross-modal HMAC-SHA-256 binding that blocks cross-patient splicing attacks per-file integrity checks cannot detect.
  • Designed a novel pre-encryption uncertainty quantification stage comparing Dempster-Shafer Theory vs Deep Evidential Learning head-to-head (Shannon entropy, size, and magic-byte features), cryptographically embedded in the bundle manifest so receivers know source-data trustworthiness at decryption time (3ms overhead).
  • Formalized security under a Dolev-Yao adversary model with a 2-256 binding-hash invariant; empirically validated 100% defense across 7 attack vectors (bit flip, intra/cross-bundle swap, truncation, injection, manifest forgery, key mismatch); RSA-OAEP hybrid key encapsulation for out-of-band key transit.
MARI — Multi-Axis Risk Intelligence Fraud Decision Engine mari-alpha.vercel.app | GitHub
Python · FastAPI · XGBoost Ensemble · Isolation Forest · SVM · Dempster-Shafer · SHAP · Docker · Railway · React + TypeScript
  • Architected a 2D Risk × Uncertainty routing matrix over three independent axes — 5-model bootstrap XGBoost ensemble risk, ensemble disagreement (epistemic uncertainty), and Isolation Forest novelty — producing 5 actionable states (APPROVE, STEP_UP_AUTH, ESCALATE_INVEST, DECLINE, ABSTAIN) instead of a single fraud score; ROC-AUC 0.9851 on 284,807 transactions.
  • Progressively eliminated human review across 4 versions: SVM second-opinion in PCA space (clears 16.07% of abstentions), Dempster-Shafer evidence fusion with conflict-factor thresholding (automates 44.44% of escalations), and SHAP-backed structured deferral — reaching a 100% automated decision lifecycle with zero false auto-blocks on legitimate transactions.
  • Integrated cost-aware decisioning (expected-loss and net-utility returned per transaction) and isotonic calibration (74.4% Brier score reduction); deployed as a FastAPI middleware API on Railway with a React dashboard on Vercel.
Technical Skills
Backend & Systems FastAPI, Node.js, Express, REST API Design, Celery, Redis, Docker, Docker Compose, Async I/O, SQLAlchemy 2.0 Databases PostgreSQL, PostGIS, Redis, SQLite, Supabase Languages Python, Java, C/C++, SQL, TypeScript, Dart ML & CV XGBoost, Scikit-Learn, SHAP, Isolation Forest, Dempster-Shafer Theory, Deep Evidential Learning, Tesseract, EasyOCR Frameworks React, Next.js, Flutter, Electron Security AES-256-GCM, HMAC, RSA-OAEP, JWT (RS256), Threat Modeling (Dolev-Yao) DevOps / Tools Git, GitHub Actions, CI/CD, Linux, Railway, Vercel, ReportLab
Publications & Certifications