Bodruddoza Araf

Full-Stack Developer (Contract), Jack The Jelli

June 2026 - August 2026. Leather goods brand, Dhaka.

• Shipped a complete e-commerce platform as the sole engineer in two months, covering storefront, cart, cash-on-delivery checkout, order tracking, customer accounts and an admin back-office in a single Next.js 16, TypeScript and MongoDB codebase.
• Designed order placement to stay correct under failure: idempotency keys block duplicate orders, guarded atomic decrements restore stock exactly once on cancellation, and snapshotted line items stop later product edits rewriting order history.
• Hardened a public, unauthenticated checkout with rate-limited server actions and server-side delivery pricing across all 64 districts, plus guest orders that attach to an account on email verification.
• Cut launch infrastructure cost to effectively zero by persuading the founders to defer the payment gateway, AI search and third-party APIs until after scale.

EduBridge AI: AI Study-Abroad Advisor

Next.js 16, TypeScript, PostgreSQL, Gemini API, Clerk, Zod, Neon, Drizzle ORM

GitHub, Live

• Built a university and scholarship matching platform for first-generation international students to a live client brief from Freelancer.com, pairing deterministic eligibility scoring with the Google Gemini API so every recommendation carries an explanation, strengths and a roadmap.
• Implemented Clerk authentication with webhook user sync, Zod-validated intake, a conversational AI advisor and a recommendations dashboard on Neon Postgres through Drizzle ORM.

PC-Builders: PC Configurator and Storefront

React 19, Node.js, Express, MongoDB, Redis, Vite, Zustand, JWT, Cloudinary, bcrypt

GitHub, Live

• Built a full PC build configurator spanning CPU, motherboard, RAM, GPU and storage as one coherent specification, plus a budget-constrained selector that assembles a compatible component set to a price the customer picks.
• Implemented the backend with JWT auth, Redis caching, Cloudinary storage and bcrypt hashing, behind a deployed React 19 and Vite frontend using Zustand for state.

News Topic Classification: NLP Architecture Comparison

Python, TensorFlow/Keras, Gensim, scikit-learn

• Ran a controlled comparison, as sole implementer, of three preprocessing strategies, two word representations (TF-IDF, custom Skip-gram) and six architectures (DNN, SimpleRNN, GRU, LSTM, Bi-GRU, Bi-LSTM), reaching 91.9% accuracy and 0.919 macro F1.
• Recovered two dead baselines: diagnosed exploding-gradient collapse at chance accuracy (LSTM 25.0%, SimpleRNN 26.6%) and brought them to 91.4% and 87.3% with gradient clipping and targeted regularisation.
• Drove preprocessing from exploratory analysis, using bigram frequency to isolate HTML noise and POS tagging to justify lemmatisation over stemming, and removed 22,067 duplicate records.

Deep Learning-Based Complete Image Prediction from Partial or Occluded Images

Undergraduate thesis, defended 2026. Supervisor: Dr. Chowdhury Mofizur Rahman, BRAC University. PyTorch.

• Co-developed a GAN-free, single-pass image completion framework (Dual-Manifold VQ-VAE, Latent Attention U-Net Mapper and Multi-Scale Detail Injection) reaching 27.60 dB PSNR and 0.861 SSIM at roughly 32 ms per 256x256 image.
• Owned mask generation, zero-shot evaluation on four unseen datasets, retraining of five baselines under matched conditions, and the ablations showing a 52.7% cut in mapper validation latent loss; placed 2nd of 6 on zero-shot average PSNR.