Skip to content
Aditya Shukla

Projects

Everything I've built. Two have full write-ups, the rest link straight to GitHub or their live links.

Pinned

Detection output from the drill rig detector across held-out validation frames
Detection output across held-out validation frames

Ternary QAT YOLO26s drill rig detector

Small-object detection with ternary QAT and a custom C++ inference runtime

  • 15,311 labeled images, single-class drill rig detection
  • 43 targeted 1x1 convolutions ternarized across the YOLO26s backbone, neck, and attention blocks
  • Per-layer sensitivity profiling, mAP50-95 0.95216 to 0.95207 at one ternary layer, 0.91056 at ten, decided which layers were safe
  • ~0.2-0.25 bytes per ternary weight via 5-trit packing, vs. 4 bytes in FP32
  • Custom AVX2 + OpenMP C++ ONNX Runtime operator, ~3.28s to ~0.78s per inference, a ~4x speedup
  • Python
  • PyTorch
  • Ultralytics YOLO
  • C++17
  • ONNX Runtime
  • CMake
Read more about the project
The PatchNote landing page, showing a generated changelog with version entries
PatchNote's landing page, with a real AI-generated changelog preview

PatchNote

Turns GitHub and Linear activity into release notes automatically

  • Integrates GitHub + Linear via OAuth 2.0
  • Gemini API converts pull requests and issues into structured release notes
  • Structured for technical, executive, and customer audiences
  • Manual effort down from ~40 minutes to under 5 minutes
  • Shareable public release pages
  • Next.js
  • Clerk
  • Supabase
  • Gemini API
  • Tailwind CSS
View live
The Vidyut AI landing page, an AI-powered forecasting platform for energy demand
Vidyut AI, forecasting electricity demand across 30 Indian smart cities

Vidyut AI

Real-time electricity demand forecasting for 30 Indian smart cities

  • XGBoost + Linear Regression hybrid
  • 9.12% MAPE on unseen test data
  • Lag-free feature pipeline using weather + calendar features, no historical demand inputs needed
  • FastAPI backend + Next.js dashboard
  • <300ms end-to-end prediction latency
  • Python
  • XGBoost
  • Scikit-learn
  • FastAPI
  • Next.js
  • Tailwind CSS
Read more about the project
ShellPDF's desktop interface with the conversion queue and status panel
ShellPDF's CustomTkinter interface, ready to queue a batch conversion

ShellPDF

Windows batch converter for Office/LibreOffice documents to PDF

  • Native MS Office COM automation with headless LibreOffice CLI fallback
  • Multithreaded conversion queue with progress and ETA
  • Windows Explorer right-click context menu integration
  • Single-instance IPC architecture
  • MIT licensed, 2 GitHub stars
  • Python
  • CustomTkinter
  • Windows COM Automation
  • LibreOffice CLI
View on GitHub

Everything

PreviewProjectTagLink
Detection output from the drill rig detector across held-out validation frames
Ternary QAT YOLO26s drill rig detectorComputer VisionRead more →
The Vidyut AI landing page, an AI-powered forecasting platform for energy demand
Vidyut AIMachine LearningRead more →
ShellPDF's desktop interface with the conversion queue and status panel
ShellPDFDesktop ToolGitHub
The PatchNote landing page, showing a generated changelog with version entries
PatchNoteFull-StackLive
Aletheia's library view, listing papers with ready, processing, and failed states
AletheiaRAGLive
Scout's chat landing page with suggested World Cup 2026 questions
ScoutAgentsLive
job-tracker's overview dashboard with live match and monitoring counts
job-trackerFull-StackPrivate repo
The RAG chatbot answering a question with a document-grounded response
RAG Contextual ChatbotRAGPrivate repo