ASHWATH SHIVA [01 // PROJECTS]
Resume PDF
Return to Portfolio
[DOCUMENT // 03] Comprehensive Technical Archive

All Projects.

A complete technical archive of implemented systems, machine learning models, and web applications. Every project details authentic problems, architectural decisions, and verified source code.

CASE STUDY 01 // RAG ARCHITECTURE
In Development · Active Alpha

KTU AI

RAG-Based Intelligent Academic Assistant for KTU Students

An intelligent academic assistant engineered to navigate fragmented KTU curriculum and ordinance PDFs, providing students with immediate, syllabus-grounded answers.

Technical Implementation: Built an asynchronous ingestion pipeline that parses curriculum PDFs into semantic chunks indexed in Qdrant Cloud. When queried, high-dimensional vector search retrieves relevant passages via a Python FastAPI backend, returning synthesized answers with verifiable PDF citations.

  • Semantic Vector Search: High-dimensional embeddings in Qdrant Cloud for syllabus module retrieval
  • Verifiable Citations: Exact ordinance, chapter, and section citation mapping for every synthesized response
  • Asynchronous Backend: High-concurrency FastAPI query endpoints connected to PostgreSQL / Supabase storage
  • Interactive Inspector: React and TypeScript interface featuring real-time query and retrieval inspection
  • React
  • TypeScript
  • FastAPI
  • Python
  • Qdrant Cloud
  • Supabase / PostgreSQL
  • RAG Pipeline
Inspect Repository
Monolithic dark obsidian and basalt archival document slabs sliced by a precision vermilion laser beam representing semantic retrieval
KTU-RAG // PIPELINE QUERY INSPECTOR SIMILARITY: COSINE >= 0.88
$ query: "What are the module prerequisites for CST 306 Algorithm Analysis?"
$ retrieval: Matched 3 chunks in qdrant_ktu_syllabus
Source: KTU_CS_S6_Ordinance.pdf § 4.2 (p. 18)
$ synthesis: CST 306 requires completion of CST 201 (Data Structures). Focus modules cover Divide-and-Conquer, Dynamic Programming, Greedy Algorithms, and NP-Completeness.
CASE STUDY 02 // AUDIO SIGNAL PROCESSING & ML
Completed Machine Learning Project
Sculptural 3D acoustic spectrogram topology carved into dark basalt stone with a vermilion contour line tracing frequency peaks in an anechoic chamber
SIGNAL PIPELINE // LIBROSA AUDIO DECOMPOSITION 40-BAND MFCC
Features Extracted: Mel-Frequency Cepstral Coefficients + Spectral Centroid + Zero-Crossing Rate
Acoustic Classes: Siren, Gunshot, Drilling, Dog Bark, Engine Idling, Street Music, Children Playing

Urban Sound Classification

Deep Learning Environmental Acoustic Event Recognition

An audio machine learning system that accurately identifies and classifies complex environmental sounds in noisy urban acoustic scenes.

Technical Implementation: Converted raw 1D audio waveforms into 2D spatial frequency representations using Librosa to extract 40-band Mel-Frequency Cepstral Coefficients (MFCCs), spectral centroids, and zero-crossing rates. Trained a custom 2D Convolutional Neural Network (CNN) in PyTorch incorporating alternating convolution layers, batch normalization, max pooling, and dropout regularization.

  • Acoustic Decomposition: Extracted 40-band MFCCs and spectrogram representations from raw audio signals
  • 2D CNN Architecture: Built in PyTorch with ReLU activations, batch norm, and spatial max pooling
  • Overfitting Regularization: Applied dropout to reliably generalize across high-variance ambient audio
  • 10-Class Recognition: Categorizes sirens, street music, drilling, engine idling, gunshots, and barking
  • Python
  • PyTorch
  • Librosa
  • NumPy
  • scikit-learn
  • Matplotlib
Inspect Repository
CASE STUDY 03 // HACKATHON PROJECT
Hackathon Project · TinkerHub Useless Projects 3.0

WHO_CARES?

Satirical Human-Evaluation & Social Ranking Platform

An unconventional human-evaluation platform that turns everyday interactions into a satirical experience. Create profiles, discover people, rate different traits, explore global rankings, and generate entertaining evaluation reports through an interactive web application.

Technical Implementation: Developed as a collaborative hackathon project for TinkerHub Useless Projects 3.0 using React, JavaScript, Vite, and Tailwind CSS. Built dynamic client-side personality dossier generation, an interactive multi-trait evaluation matrix, QR encounter verification, and randomized 50/50 appeal resolution with Lucide iconography and cryptographic integrity via the Web Crypto API.

  • Profile Dossier Management: Create and manage dynamic profiles with humorous traits and customized persona metrics
  • Human Discovery & Random Encounters: Discover community members through random summoning or QR-based encounter codes
  • Interactive Trait Scoring: Granular multi-factor rating system evaluating satirical behavioral parameters
  • Global Leaderboard & Tiers: Dynamic rankings categorizing evaluated profiles into competitive status tiers
  • Evaluation Reports & Certificates: Exportable Humanity Certificates and detailed satirical evaluation dossiers
  • Coin-Toss Appeal Court: Built-in 50/50 randomized appeal mechanism allowing users to contest unfavorable ratings
  • React
  • JavaScript
  • Vite
  • Tailwind CSS
  • Web Crypto API
  • Lucide Icons
Inspect Repository
Faceted obsidian sculptural bust on a dark basalt pedestal with suspended metric planes and subtle vermilion laser accents representing satirical human evaluation
EVALUATION DOSSIER // TINKERHUB USELESS PROJECTS 3.0 TIER: S-RANK CITIZEN
Evaluation Metrics: Aura Index: +420 · NPC Likelihood: 12% · Zombie Survival: Low · Reply Latency: 3.2d
Encounter Verification: Cryptographic Humanity Hash via Web Crypto API · Appeal Mechanism: 50/50 Coin Toss
CASE STUDY 04 // FULL STACK WEB APPLICATION
Completed Web Application

Period Tracker

Private Cycle Tracking & Health Calendar Web Application

A lightweight, privacy-focused web application for menstrual cycle tracking and phase prediction with zero third-party telemetry.

Technical Implementation: Developed a responsive full-stack architecture using PHP and MySQL. Engineered custom phase prediction algorithms calculating future cycles from user-entered historical durations, stored data in a normalized relational schema with parameterized SQL queries to prevent injection, and secured access with local session authentication.

  • Predictive Calculation: Algorithmic cycle phase and ovulation estimation based on historical cycle averages
  • Interactive Calendar: Fast client-side calendar visualization and daily symptom tracking using Vanilla JS and CSS
  • Zero Telemetry: Strict privacy architecture with local session management and zero external tracker scripts
  • Mobile-First Performance: Lightweight footprint with relational MySQL schemas and parameterized queries
  • PHP
  • HTML5
  • CSS3
  • JavaScript
  • MySQL
Inspect Repository
Abstract cyclical orbital sculpture of precision-milled matte obsidian and dark graphite rings with a glowing vermilion phase arc indicator
SYSTEM METRICS // ZERO TELEMETRY CONFIRMED PHP/MYSQL STACK
Security & Privacy: Local session authentication, parameterized SQL queries, zero third-party trackers
Frontend Execution: Vanilla CSS responsive calendar grid, lightweight JavaScript date calculation engine
[RESEARCH // 05] Ongoing Exploration & Scaled Architectures

Continuous Engineering & Open Source Exploration

Currently pursuing specialized technical explorations in high-dimensional vector quantization, transformer fine-tuning on academic corpora, and asynchronous backend microservices. New verified repositories are published directly to GitHub upon reaching reproducible milestones.

Follow on GitHub Propose a Collaboration