✦ Fashion Tech Hub · Unified Launcher
Atelier — Fashion Tech Hub
A single landing page tying together my fashion and style AI projects — WYA (wardrobe intelligence), Luna (conversational stylist), TryOn (3D avatar preview), and Velaris (AI jewellery design). No shared backend or session — each app keeps its own auth, Atelier is purely the front door. Static site with a custom pink-to-teal gradient identity, deployed on GitHub Pages.
Static Site
HTML/CSS
GitHub Pages
Design System
Systems Programming · Currently Building
Lumen — A 32-bit x86 Operating System
An operating system built from a raw boot sector up: custom BIOS bootloader and a freestanding C kernel of roughly 13,000 lines. Implements paging with per-task address spaces, a bitmap frame allocator, an interrupt-safe heap, preemptive Ring 0/Ring 3 multitasking with fork/exec, a 25-call syscall table, an ATA driver, a persistent filesystem (LumenFS v2), an ELF loader, and a text-mode GUI. A make test harness boots it headless in QEMU and fails on any self-test failure.
C
x86 Assembly
NASM
QEMU
GitHub Actions
Full-Stack AI + AWS Cloud + Computer Vision · Fully Built
WYA — What's Your Aesthetic?
Full-stack wardrobe intelligence platform. Users upload garments and a two-tier pipeline — AWS SageMaker (FashionCLIP) for zero-shot classification + a local rule-based fabric classifier — auto-tags every item with category, color, fabric, and pattern. A Style DNA quiz maps aesthetic preferences to a style vector that drives outfit matching, gap analysis, and sustainability scoring. Production FastAPI backend on AWS EC2 behind CloudFront, rate-limited AI endpoints, JWT auth, and 47 automated tests.
React
TypeScript
FastAPI
Python
Docker
FashionCLIP · SageMaker
AWS EC2 · S3 · CloudFront
GitHub Actions CI/CD
SQLite
Conversational AI · WYA Integration · Fully Built
Luna — Conversational AI Stylist
A chat interface that lets you talk to your wardrobe in plain English. A regex-based intent classifier routes structured queries — outfit suggestions, wardrobe search, gap analysis, style explanation — to the right WYA API endpoint, while open-ended fashion questions fall back to an OpenRouter LLM chain with the user's real wardrobe passed in as context. Own FastAPI server (Docker, EC2), separate CloudFront deployment, JWT auth shared with WYA.
React
TypeScript
FastAPI
Powered by WYA API
Intent Classifier
OpenRouter LLM Fallback
AWS EC2 · S3 · CloudFront
Docker
Machine Learning · FinTech
Bitcoin Fraud Detection
Advanced fraud detection using Graph Neural Networks (GATv2) on the Elliptic Bitcoin dataset. Features a MetaEnsemble stacker, heterophily-aware readout, NT-Xent contrastive loss, and exceeds published EvolveGCN baselines on AUC and F1.
GNN · GATv2
Python
PyTorch Geometric
XGBoost
Computer Vision · Fully Built
TryOn — Build, Compare, and Perfect Outfits
Turns any clothing screenshot — Pinterest, Instagram, a store listing — into a transparent garment cutout via rembg, auto-categorizes it with a FashionCLIP zero-shot classifier ported from WYA, and fits it onto a 2D paper-doll or a rigged 3D mannequin using MediaPipe BlazePose landmarks and thin-plate-spline warping. Users build, save, and compare full outfit combinations, then export a finished look as a shareable magazine-style card. FastAPI backend with 69 tests at 86% coverage, run against the real CV models rather than mocks.
React
TypeScript
FastAPI
rembg · MediaPipe
FashionCLIP
React Three Fiber
SQLite
Computer Vision · Part of TryOn
3D Mannequin Garment Viewer
Upload a single front-facing garment photo and a deterministic CV analyzer detects strap type and back construction, then renders the result on an interactive 3D dress-form mannequin with a matching generated back texture. An anti-hallucination gate reports "undetermined" below 75% confidence instead of inventing a back. Runs against an optional FastAPI backend and falls back to an equivalent client-side Canvas engine when it's offline.
React
TypeScript
Three.js
FastAPI
Pillow · Scanline CV
Tailwind CSS
Reinforcement Learning · Clean Energy
HydroDrive-DQN — Adaptive FCEV–Hydrogen Co-Optimization
Resilience-aware pipeline coupling fuel-cell EV energy management with hydrogen infrastructure planning. NSGA-II sets system-level priorities, a from-scratch NumPy DQN allocates fuel-cell vs battery power using live hydrogen and renewable context, a genetic algorithm sizes the supply chain, and a 10,000-scenario Monte Carlo run stress-tests demand satisfaction. Benchmarked against rule-based, thermostat, standard-DQN, and MILP-style baselines across WLTP, NEDC, US06, and UDDS cycles, with 17 auto-generated figures and a Flask dashboard.
Deep Q-Network
NSGA-II
Genetic Algorithm
Monte Carlo
Python · NumPy
Flask
Causal ML · Fairness in Deepfake Detection
FairFaceGuard — Deepfake Detection Under Skin-Tone and Lighting Variations
Research pipeline that asks whether a deepfake detector's accuracy gap across skin-tone groups is really about skin tone, or about illumination cues that merely correlate with it. Faces get illuminant-corrected ITA/Fitzpatrick labels, then matched counterfactuals (skin-only, light-only, both) with an automated manipulation check. A frozen EfficientNet-B4 baseline is interrogated three ways — adversarial GRL sweeps, linear probing, and a counterfactual causal test — with bootstrap CIs and a paired permutation test. Currently a single-seed pilot on a FaceForensics++ Deepfakes-only subset; multi-seed, cross-dataset (Celeb-DF) and backbone-ablation runs are written but not yet run.
Causal Counterfactuals
Python
PyTorch
EfficientNet-B4 · FaceForensics++
GRL Disentanglement
Bootstrap CI · Permutation Test
Multimodal AI · Hallucination Mitigation
VisualGuard — Evidence-Guided Decoding for LVLMs
Training-free, decoding-time method that cuts object hallucination in Large Vision-Language Models without touching the weights. At every step it scores each top-k candidate token by how well the image supports it, using LVLM attention, CLIP similarity and optional Grounding DINO region detections, then penalises the logits of unsupported candidates before the next token is chosen. Includes a KV-cached decoding loop, a per-step intervention audit trail, POPE and MME evaluation on LLaVA-1.5, ablation runners, and 297 tests that need no downloads. Full benchmark results have not been run yet.
LVLM · Hallucination Reduction
Python
PyTorch
Transformers
CLIP · Grounding DINO
LLaVA-1.5
POPE · MME
Deep Learning · Motorsport Analytics
ApexNet — F1 Overtaking Prediction
Predicts whether the driver starting behind a grid-adjacent rival finishes ahead. Trains and compares five architectures (CNN, GAT, LSTM, CNN+LSTM, and a proposed CNN+GAT) on real f1db results from 2014 onward, with a leakage-checked race-level split, held-out confusion matrix, and feature-group ablation. A live replay dashboard checks predictions against what actually happened — no synthetic telemetry anywhere in the training pipeline.
CNN + GAT
PyTorch
FastAPI
scikit-learn
Pandas
Data Visualization · Sports Analytics
Sports Performance Analytics Dashboard
Six-tab live dashboard covering F1, MotoGP, tennis, Real Madrid, San Jose Sharks, and equestrian, polling free public APIs straight from the browser with no backend and a freshness badge on every tab. Includes a Dixon-Coles-style Poisson match model validated walk-forward on 14,460 real matches across four leagues, with negative results reported honestly. 160 tests, including a leakage canary on lagged form features.
React 19
TypeScript
Tailwind CSS
Recharts
Vite
Poisson Model
Generative AI · Currently Building
Velaris — From Inspiration to Jewellery Design
An AI-powered jewellery design tool that turns a text description, a rough sketch, or an inspiration photo into a professional concept render, a structured specification sheet (type, metal, stone, cut, style, setting, occasion), and an exportable PDF design brief a jeweller can quote and manufacture from. Built to close the gap between a customer's vague idea and a jeweller's workflow — cutting down the back-and-forth of draft revisions before a single piece is made.
React
Generative AI · Image Generation
Spec Sheet Generation
PDF Export
NLP Pipeline · Text Style Transfer
FlowWrite — NLP Text Style Transfer Pipeline
Multi-pass NLP pipeline that rewrites text to match natural human writing patterns. Combines spaCy chunking, SBERT semantic similarity for candidate ranking, and a local LLM (llama3.2 via Ollama) for rewriting. A custom Human Likeness Score (HLS) evaluates outputs across burstiness, readability, coherence, and semantic preservation.
NLP Pipeline
Python
FastAPI
spaCy
SBERT
Ollama · llama3.2
AWS S3 · CloudFront
Full-Stack
BookNest+
Smart personalized reading platform to discover, track, and manage books — powered by mood, behavior, and reading analytics. Features a custom DRPA recommendation engine, EPUB reader, book-to-movie mode via TMDB.
TypeScript
React
Node.js
Docker
Recommendation AI
Machine Learning
Smart Resale
ML model predicting fair resale prices for used phones and laptops. Covers data preprocessing, feature engineering, and model evaluation with Pandas, NumPy, and Scikit-learn.
Machine Learning
Python
Scikit-learn