ria = {
"studying": "CS & Data Science @ MUJ (3rd year, 2024–2028)",
"currently_building": ["Velaris", "Lumen", "PaperPilot", "Atelier"],
"shipped": ["WYA", "Luna", "TryOn"],
"accepted": ["WYA @ EDGEI 2026 (Springer LNNS)",
"FairFaceGuard @ IMPACT 2026 (Springer ICT)"],
"ip": ["WYA — IDF Accepted (Manipal University Jaipur, 2026)"],
"interests": ["Full-Stack Dev", "Machine Learning", "Graph ML",
"Generative AI", "Fairness & Causal ML", "Research"],
"stack": ["Python", "TypeScript", "React", "FastAPI", "PyTorch", "AWS"],
"looking_for": "Summer 2027 internships — SWE · ML · Data Science (remote / India)",
"fun_fact": "I build apps to solve problems I actually have"
}| 📄 2 papers accepted | WYA at EDGEI 2026 (Springer LNNS, Scopus-indexed) · FairFaceGuard at IMPACT 2026 (Springer ICT series) |
| 📤 4 papers submitted | TryOn (ANTIC 2026) · VisualGuard (ICIPCN 2027) · HydroDrive-DQN (MARC 2027) · Velaris (ICDSI 2027) |
| 📜 IP | WYA IDF accepted · TryOn and Greywater patent drafts submitted |
| 🚢 Shipped | Production AWS stack (EC2 · S3 · CloudFront · SageMaker) with CI/CD and automated tests |
Ever felt like your style exists but you just can't name it?
🏆 Paper accepted at EDGEI 2026 (camera-ready submitted) · 📜 IDF Accepted · Manipal University Jaipur, 2026
WYA is a full-stack AI wardrobe intelligence platform. Upload garments and a two-tier pipeline (AWS SageMaker FashionCLIP + a local fabric classifier) auto-tags category, colour, fabric and pattern. A Style DNA quiz builds your aesthetic profile, then drives outfit matching, wardrobe gap analysis and sustainability scoring.
✦ FastAPI backend on AWS EC2 + CloudFront ✦ FashionCLIP zero-shot tagging (SageMaker)
✦ JWT auth + rate-limited AI endpoints ✦ Style DNA engine
✦ Gap analysis with affiliate mapping ✦ Outfit generation
✦ Sustainability scoring ✦ 47 automated tests · GitHub Actions CI/CD
ℹ️ The AWS backend is currently paused for cost management.
Talk to your wardrobe in plain English.
A conversational frontend on top of WYA's API. A regex intent classifier routes structured queries (outfits, wardrobe search, gap analysis, style explanation) to the right WYA endpoint; open-ended fashion questions fall back to an LLM chain with your real wardrobe as context.
✦ Natural-language intent classifier ✦ LLM fallback for any fashion question
✦ Routes to WYA's real API ✦ React + TypeScript
✦ Own FastAPI server (Docker, EC2) ✦ AWS S3 + CloudFront · JWT auth shared with WYA
Screenshot a look. Wear it on a mannequin.
Turns any clothing screenshot into a transparent garment cutout (rembg), auto-categorises it with FashionCLIP, and fits it onto a 2D paper-doll or rigged 3D mannequin using MediaPipe BlazePose landmarks and thin-plate-spline warping. Build, save and compare full outfits, then export a magazine-style card. 69 tests at 86% coverage, run against the real CV models.
🔗 View Project → · 📤 Paper submitted · ANTIC 2026 · 📝 IDF filed
Is the accuracy gap really about skin tone, or about the lighting that comes with it?
🏆 Paper accepted at IMPACT 2026 · Springer Information and Communication Technology (ICT) series
A research pipeline that audits a deepfake detector across skin-tone groups. 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. The accepted paper presents an observational fairness audit together with an honest negative result on counterfactual skin-tone control.
✦ Illuminant-corrected ITA / Fitzpatrick labels ✦ Matched counterfactuals + manipulation check
✦ Adversarial GRL sweeps · linear probing ✦ Bootstrap CIs · paired permutation test
✦ Frozen EfficientNet-B4 on FaceForensics++ ✦ Negative results reported honestly
Type it. Sketch it. Upload it. Get a jewellery design you can actually build.
📤 Paper submitted · ICDSI 2027 (Paper ID 391)
An AI-powered jewellery design tool that turns text, sketches or inspiration photos into professional concept renders and manufacturer-ready briefs — closing the gap between a customer's idea and a jeweller's workflow.
✦ Text / sketch / photo input ✦ AI concept image generation
✦ Structured specification sheets ✦ Exportable PDF design briefs
✦ Manufacturer-ready output ✦ Planned: multi-view & cost estimation
Core flow: input (text / sketch / photo) → AI concept image → spec sheet (type, metal, stone, cut, style, setting, occasion) → PDF for jeweller quotation.
| Project | Description | Tech |
|---|---|---|
| Lumen | 32-bit x86 OS from a raw boot sector: paging, preemptive Ring 0/3 multitasking, fork/exec, 25 syscalls, ATA driver, persistent FS, ELF loader |
C · x86 ASM · NASM · QEMU |
| VisualGuard | Training-free, decoding-time method that reduces object hallucination in LVLMs by scoring candidate tokens against image evidence · 297 tests | PyTorch · CLIP · Grounding DINO · LLaVA-1.5 |
| HydroDrive-DQN | Fuel-cell EV energy management + hydrogen infrastructure co-optimization: from-scratch NumPy DQN, NSGA-II, GA, 10,000-scenario Monte Carlo | Python · NumPy · Flask |
| ₿ Bitcoin Fraud Detection | GATv2 + MetaEnsemble stacker on the Elliptic dataset; beats published EvolveGCN baselines on AUC & F1 | PyTorch Geometric · XGBoost |
| ApexNet | F1 overtaking prediction — five architectures compared (CNN, GAT, LSTM, CNN+LSTM, CNN+GAT) with a leakage-checked race-level split | PyTorch · FastAPI |
| Sports Analytics Dashboard | Six-tab live dashboard + a Poisson match model validated walk-forward on 14,460 real matches · 160 tests | React 19 · TypeScript · Recharts |
| PaperPilot | Questions in, handwritten A4 pages out — custom canvas renderer + hand-written PDF writer, answers streamed from Claude over SSE | JavaScript · FastAPI · Redis |
| flowWrite | Multi-pass NLP pipeline for natural-sounding rewrites with a custom Human Likeness Score | spaCy · SBERT · Ollama |
| 3D Mannequin Viewer | Garment photo → strap/back detection → 3D dress-form render, with an anti-hallucination confidence gate | Three.js · FastAPI |
| BookNest+ | Mood- and behaviour-driven book recommendations, EPUB reader, book-to-movie mode | TypeScript · React · Node.js |
| Smart Resale | ML model predicting fair resale prices for used phones & laptops | Scikit-learn · Pandas |
| Title | Type | Ref. No. | Status |
|---|---|---|---|
| WYA — What's Your Aesthetic: AI-Powered Personal Styling and Wardrobe Optimization | Invention Disclosure Form | TEMP/E-1/134770/2026-DEL | ✅ IDF Accepted · fee paid, filing in process |
| TryOn — Adaptive Warping Method for Garment Fitting on Virtual Mannequins | Patent draft | — | 📝 Draft submitted |
| Greywater | Patent draft | — | 📝 Draft submitted |
Co-inventor with Mitakshi Sinha · Manipal University Jaipur, 2026
- 🏆 Two papers accepted: WYA at EDGEI 2026 and FairFaceGuard at IMPACT 2026 (camera-ready in progress) · 📜 WYA IDF accepted at Manipal University Jaipur
- 📤 Papers under review: TryOn, VisualGuard, HydroDrive-DQN, Velaris
- 🚧 Writing up the Bitcoin GNN and WYA journal papers (The Computer Journal, OUP)
- 💎 Building Velaris · Lumen OS · PaperPilot · ✦ Atelier (fashion-tech hub tying WYA, Luna, TryOn & Velaris together)
- 🎯 Deepening ML: GNNs, causal fairness, LVLM decoding, reinforcement learning
- 🌐 Exploring backend architecture, API design & cloud infrastructure