Selected Projects & Systems
A curated index of production-grade Machine Learning architectures, deep learning research systems, and public web platforms engineered with algorithmic precision, modularity, and high performance.
Machine Learning & Research Systems
Deep learning architectures, multimodal transformer encoders, and automated MLOps pipelines engineered for empirical rigor, reproducible benchmarks, and scalable deployment.
-
01 A Unified Multimodal Fake News Detection Framework (BERT + ViT)
90.18% Test Acc 0.9485 ROC-AUC Springer IC2NS2 '26Deep Learning • Multimodal Transformers • Research Publication- Core Architecture: End-to-end multimodal classification pipeline integrating BERT–BiGRU for contextual narrative semantics with a Vision Transformer (ViT-B/16) for patch-level visual representations.
- Empirical Outcome: Employs efficient feature concatenation fusion to achieve 90.18% Test Accuracy, 0.9485 ROC-AUC, and 0.9084 Macro F1 across 30,900 samples on the Fakeddit benchmark.
-
02 Coccidiosis Chicken Disease Classification (Deep Learning + MLOps)
VGG16 Transfer Learning AWS EC2 Deployment Docker + DVCComputer Vision • Transfer Learning • MLOps Pipeline • CI/CD Automation- Core Architecture: Production-grade deep learning classification system engineered for automated poultry disease diagnosis (Coccidiosis) utilizing fine-tuned VGG16 transfer learning backbones with modular data ingestion and validation components.
- MLOps Infrastructure: Integrated DVC (Data Version Control) for reproducible data/model artifact versioning, MLflow for experiment tracking and metric evaluation, containerized with Docker & FastAPI, and deployed autonomously to AWS EC2 via GitHub Actions CI/CD.
-
03 End-to-End Machine Learning & MLOps Student Performance Pipeline
Live Streamlit App CatBoost & XGBoost MLflow + SHAPMachine Learning • MLOps • Model Interpretability (SHAP) • Streamlit Deployment- Predictive Modeling Architecture: Modular regression engineering pipeline predicting student mathematical performance by systematically training, tuning, and evaluating multiple models including Ridge, Random Forest, XGBoost, and CatBoost.
- Interpretability & Live Deployment: Features SHAP (SHapley Additive exPlanations) for global and local feature attribution to demystify socioeconomic score drivers, tracked through MLflow, and served as an interactive real-time Streamlit application.
-
04 Academic Research & Engineering Architecture Hub
Astro 5 SSG Design Tokens Zero-BloatModern Web Architecture • Academic Portfolio • Accessible Typography • PWA- Algorithmic Portfolio Platform: High-performance, typography-focused academic portfolio and research repository built with Astro SSG, structured data schemas (JSON-LD), and custom CSS design tokens.
- Interactive Systems: Integrates interactive SVG deep learning architecture visualizers, instant BibTeX citation export engines, responsive CSS Grid layouts, and light/dark theme persistence with zero heavyweight client runtime overhead.
Production Web Platforms & Public Utilities
High-performance, privacy-first consumer platforms, deep learning hardware analytics engines, and specialized mathematical utilities engineered for zero-latency client execution, deep localized SEO, and global reach.
-
01 GPUCalc — Deep Learning & LLM VRAM Sizing Engine
Live Web Platform Cloudflare Workers LLM ParallelismDeep Learning Infrastructure • VRAM Memory Analytics • Serverless Cloudflare EnginePrecision: FP32 / FP16 / BF16 / FP8 / INT4 (AWQ/GPTQ) Parallelism: TP • PP • DP • ZeRO 1-3 • FSDP Runtime: Cloudflare Workers Edge (Sub-15ms TTFB)- Analytical Memory Architecture: Hardware calculation engine estimating exact GPU VRAM requirements and bandwidth throughput boundaries for LLM pre-training, fine-tuning, and inference workloads across diverse precision regimes (FP16, BF16, INT8, FP8, and 4-bit AWQ/GPTQ/GGUF).
- Distributed Topology & ZeRO Analysis: Granular memory modeling for optimizer state overhead (16 bytes/param for AdamW), gradient staging, and activation checkpointing across DeepSpeed ZeRO-1/2/3 and PyTorch FSDP sharded configurations.
- Inference KV-Cache & Edge Runtime: Dynamically calculates attention cache scaling (2×b×s×l×h×d) supporting MHA, GQA, and MLA across NVIDIA H100, A100, and RTX 4090 GPUs, deployed on Cloudflare Workers for sub-15ms execution.
-
02 AllAgeCalculators.com — High-Precision Chronological & Date Computation Engine
Live Web Platform Astro + TypeScript 100% Client-Side PWA & OfflineWeb Platform • Chronological Mathematics • Competitive Exam Matrix • Pediatric & Veterinary Analytics- Live Chronological Breakdown & Exam Eligibility: Real-time chronological engine computing exact lived age in years, months, days, minutes, and seconds alongside an algorithmic compliance matrix for Indian competitive exams (UPSC CSE, SSC CGL, IBPS PO, NDA) validating official recruitment cutoffs and legal age relaxations.
- Specialized Pediatric & Veterinary Math: Features an obstetric gestational and corrected age calculator for preterm infants, a leap-year-aware workday interval counter, and modern non-linear canine age calculations based on veterinary breed weight classifications.
- Architecture & Zero-Knowledge Privacy: Built using Astro, TypeScript, and TailwindCSS with PWA offline support. Executes 100% of mathematical routines locally in the browser with zero server data collection.
-
03 CalculadoradeIdade.com — Lusophone Chronological & Public Examination Platform
Live Web Platform Lusophone Regionalization LGPD Compliant Concursos & INSSWeb Platform • Internationalization (pt-BR / pt-PT) • Concursos Públicos • Aposentadoria INSS- Lusophone Regional Chronological Analytics: High-performance chronological engine purpose-engineered for Portuguese-speaking markets (Brazil, Portugal, Angola, Mozambique), delivering real-time chronological breakdown, milestone dates, and astrological tracking.
- Concursos Públicos & Previdência (INSS): Tailored compliance engine for Brazilian civil service examinations validating candidate age eligibility against strict edital publication dates, coupled with an INSS retirement simulator under Brazil's Pension Reform (EC 103/2019) and gestational health tools.
- Architecture & LGPD Compliance: Implemented with Astro, TypeScript, and multi-region
hreflangindexing. Operates completely in-browser with zero tracking in full compliance with Brazil's General Data Protection Law (LGPD).
-
04 PDFAfy — Fast, Privacy-First PDF to PDF/A Converter
Live Web Platform PDF/A-1b, 2b, 3b Zero Server UploadWeb Platform • Client-Side Document Processing • PDF/A Archival Standards • Zero Data ExposureStandards: PDF/A-1b • PDF/A-2b • PDF/A-3b (ISO 19005) Execution: 100% In-Browser Web Workers (Zero-Upload)- Archival Compliance Engine: Fast, in-browser PDF to PDF/A converter with full support for PDF/A-1b, PDF/A-2b, and PDF/A-3b ISO standards, facilitating document archiving, institutional submissions, and legal compliance.
- Zero-Knowledge Privacy: Operates entirely on the client side via pdf-lib and Web Workers. No sensitive files or data are ever transmitted to external servers or stored in remote caches, guaranteeing total document privacy.
- Color Profile & Web Worker Architecture: Integrates embedded sRGB color profiles, standardized XMP metadata harmonization, and offloads linearization to background workers to prevent UI lockup during large document conversions.
