Bangladesh University of Engineering and Technology

Received the Dean’s List Award and University Merit Scholarship for academic performance.

Notable courses:

• CSE 305 - Computer Architecture
• CSE 307 - Software Engineering
• CSE 309 - Compiler Design
• CSE 313 - Operating Systems
• CSE 317 - Artificial Intelligence
• CSE 321 - Computer Networks
• CSE 405 - Computer Security
• CSE 409 - Computer Graphics
• CSE 463 - Introduction to Bioinformatics

Thesis: Evaluating the Multilingual Capabilities of HTML-based Web UI Agents, supervised by Dr. Rifat Shahriyar.

Multilingual Evaluation of HTML-based Web Agents

Paper submitted to NSysS 2026 (under review)

Most web agent benchmarks are in English, but most of the web is not. We wanted to know whether HTML-based LLM agents get worse when the page is in another language.

• Fine-tuned XLM-RoBERTa and mT5 agents on Mind2Web tasks translated into seven languages (English, Bengali, Hindi, Spanish, Portuguese, Chinese, and Russian).
• Offline: built a benchmark that translates full Mind2Web pages while keeping their structure identical, so language is the only thing that changes. Even with 63% of the page’s readable text rewritten, step accuracy dropped by only about 1.5 points, mostly from the text the agent types.
• Live websites: ran a Gemini-based agent in a real browser on localized versions of sites like Google Flights, YouTube, Wikipedia, and BBC News, 270 runs in all. About 81% of what the agent reads changed between languages, yet goal completion stayed between 88.6% and 94.3%, with no significant gap against English.
• Found that the type of task mattered more than the language: tasks that need the page content read were about 16 points harder, in every language. The main cost of other languages was tokens, up to 2.4× more on non-English pages.

---

DFCon: Deepfake Face Detection

First runner-up, SP Cup 2025 · Submitted to NSysS 2026 (under review)

Deepfake detectors often do well on the data they were trained on and then fail on fakes made with methods they have not seen. Our team built a detector for the SP Cup’s “in the wild” setting.

• Combined three vision models with different strengths (MaxViT, CoAtNet, EVA-02) and fused their predictions by majority vote.
• Fine-tuned each model with supervised contrastive learning before training its classifier, and added 12,200 synthetic deepfakes from seven recent generation methods to the training data.
• Reached 95.83% validation accuracy and 0.9239 AUC on the 395,609-image hidden test set.

---

LLMSniffer: Detecting LLM-Generated Code

Preprint on arXiv

A classifier that tells whether a piece of code was written by a person or by an LLM.

• Fine-tuned GraphCodeBERT in two stages: supervised contrastive learning to pull human and LLM-written code apart in embedding space, then an MLP classifier on comment-stripped code.
• Improved accuracy over prior work from 70% to 78% on the GPTSniffer dataset and from 91% to 94.65% on Whodunit.

Shipday Inc.

Shipday is a last-mile delivery startup based in California. I work mostly on backend services, APIs, and third-party integrations.

• Built and improved integrations with delivery providers like DoorDash, Uber, and Roadie for delivery fulfillment, order status sync, and live driver tracking.
• Extended ordering and POS integrations (Toast, Shopify, Lightspeed) and moved them to a newer architecture.
• Worked on a B2C app that connects customers with third-party delivery providers for food and parcel delivery.
• Built direct-mail marketing features so customers and our sales team can send postcard and flyer campaigns.
• Built internal LLM-based sales tools: one writes structured lead summaries from business data and web search, another scores sales-call transcripts against set metrics.
• Migrated the marketing website from Webflow to Astro and Cloudflare, cutting about $12,000 in yearly subscription costs.
• Wrote a web scraper that collected over 30,000 prospective leads for sales outreach.

---

Therap (BD) Ltd.

Joined right after graduation and completed a structured training program on enterprise backend development with Java, Spring, Hibernate, Servlets, and JSP.