Dhaka, Bangladesh · Open to graduate research positions

Kiyas Mahmud

AI Researcher · Autonomous Agents & Multimodal Learning

Computer Science undergraduate at American International University-Bangladesh (AIUB). My thesis, WebFARM, studies how autonomous web agents detect failures, recover from them, and remember what went wrong. I have contributed to papers accepted at IEEE TENCON 2026 and ICECCME 2026.

Kiyas Mahmud in a grey suit
BSc CS · AIUB · Class of 2026
Publications
4

2 accepted at 2026 conferences

CGPA
3.45/4.00

BSc Computer Science, AIUB

National award
Winner

Bangladesh ICT & Innovation Awards 2025

Thesis dataset
40K

real-browser interaction steps

01 / About

Profile & education

I work where machine learning research meets software that has to run in the real world.

My research covers autonomous AI agents, multimodal learning, and explainable, trustworthy AI. Alongside it, I have led industry projects end to end, from a 540-endpoint field-sales platform for Elite Group of Industries to a multi-tenant point-of-sale system for a mobile phone wholesaler. Building systems that real users depend on shapes how I design experiments: measurable, reproducible, and honest about failure.

Research interests

  • Autonomous AI Agents
  • Multimodal Learning
  • Vision-Language Models
  • Trustworthy AI
  • Explainable AI
  • Machine Learning
  • Cloud Optimization

Education

  1. BSc in Computer ScienceDec 2026
    American International University-Bangladesh (AIUB), DhakaCGPA 3.45 / 4.00
  2. Higher Secondary Certificate, Science2021
    Ideal College, Dhanmondi, DhakaGPA 4.75 / 5.00
  3. Secondary School Certificate, Science2018
    Singair Pilot Government High School, ManikganjGPA 4.94 / 5.00

English · fluent | Bangla · native

02 / Research

Research experience

Undergraduate thesis · Lead technical researcher

WebFARM: A Multimodal Framework for Failure Detection, Adaptive Recovery and Episodic Memory in Autonomous Web Agents

Supervisor: Asst. Prof. Nuzhat Tabassum, Department of Computer Science, AIUB

  • Led the technical research: methodology and experimental design, framework architecture and implementation, model integration, experiment execution, and results analysis.
  • Developed and evaluated multimodal approaches using Qwen2-VL-2B, Qwen2.5-VL-7B, and InternVL3.5-8B on Web-Gold-40K, a dataset of 40,017 real-browser interaction steps.
  • Evaluated Browser Use and SeeAct agents across 22 websites and 42 tasks.
Best F1
0.839
Recovery prediction
93.49%
Browser Use gain
+13.1pp
SeeAct gain
+20.2pp

Preliminary results; task-completion gains measured over the unmodified agents.

Jan 2025 – Mar 2026

Research Assistant · Tech Wing's Lab, Dhaka

Assisted with research planning, analysis, documentation, and research paper preparation; reviewed academic papers and guided junior researchers in academic writing and research methods.

03 / Publications

Publications

  1. [1]

    AutoTuneX: An Intelligent Self-Optimizing Runtime Framework for Cloud Applications

    Accepted IEEE TENCON 2026, Bali, Indonesia

    Five-component closed-loop Kubernetes autoscaler. Live Minikube experiments reported 72.2% lower cost, zero SLO violations, and zero scaling oscillations using Linear Regression.

  2. [2]

    Hybrid FraudShield: Detecting Financial Fraud in Real Time with Explainable AI

    Accepted ICECCME 2026, Bali, Indonesia

    Stacking ensemble with a Logistic Regression meta-learner on about 590,000 IEEE-CIS transactions: 98.7% accuracy, 0.978 ROC-AUC, and an 8.1-point precision gain over XGBoost at equal recall, with SHAP explanations under 500 ms.

  3. [3]

    Comparative Analysis of Heap Algorithms for Anomaly Localization in Dynamic Graph Neural Networks

    Published Atlantis Press, Proceedings of ICERIE 2025 · online 18 Nov 2025

    Compared Fibonacci, D-ary, and PriorityQueue heaps in Dijkstra's Algorithm for anomaly localization, examining efficiency, memory use, and scalability.

  4. [4]

    Real-Time Fusion of Behavioral and Transactional Data for Fraud Detection

    Presented ICASF 2025, Dubai

    Ensemble framework combining behavioral, transactional, and device-level data for real-time fraud detection.

04 / Experience

Professional & industry experience

Industry collaboration

Sales Officer Copilot

Elite Group of Industries, Bangladesh

AI-powered sales management platform

Client project · private
  • Led the development of a Bangla and English sales app with GPS tracking, reporting, and live team monitoring.
  • Built a secure backend with 540 API endpoints and integrated a Gemini-powered AI sales assistant.

Python · FastAPI · Cloudflare Workers · Flutter · React · Gemini API

Client project · 2026

Phonelot

Taslima Enterprise, Bangladesh

Wholesale mobile phone POS

Client project · private
  • Developed a Bangla and English sales system for phone inventory, orders, payments, and barcode scanning.
  • Built a secure backend with inventory tracking, activity logs, backups, and 494 automated tests.

Nuxt 4 · TypeScript · Hono · Cloudflare D1 · Tesseract.js · ZXing WASM

05 / Projects

Selected projects

WICEBD Silver Medal

ScholarNest

AI-powered scholarship management platform

  • Global scholarship discovery, application tracking, professor CRM, AI-personalized emails, and document review.
  • Automated follow-ups, deadline alerts, analytics, risk scoring, and secure document management.

Nuxt 3 · Vue.js · TypeScript · PostgreSQL · Prisma ORM · Redis · Gemini AI

National ICT Award 2025 · APICTA 2025

Hybrid Fraud Shield

Explainable AI-based fraud detection

  • Hybrid ML/DL system combining XGBoost, CatBoost, CNN, and BiLSTM with ensemble learning for real-time transaction risk assessment.
  • SHAP and Integrated Gradients for explainable predictions, served through a FastAPI real-time inference service.

Python · Scikit-learn · XGBoost · CatBoost · SHAP · FastAPI

HandyLaw

AI-powered legal drafting platform

  • AI-assisted Bangla legal drafting with verified legal-source citations, structured case analysis, and mandatory advocate review and approval.
  • Secure AI authorization, Gemini output validation, sandboxed DOCX/PDF generation, and 2,400+ automated tests.

TypeScript · Hono · Cloudflare Workers · D1 · Nuxt 4 · Gemini API · Zod · Vitest

06 / Skills

Technical skills

AI / ML
Scikit-learn, XGBoost, CatBoost, SHAP/TreeSHAP, Gemini API, model evaluation, preprocessing, ensemble and stacking methods, fraud detection
Agentic systems
Autonomous web agents, failure detection, adaptive recovery, episodic memory, AI agent integration
Programming
Python, C++, JavaScript, TypeScript, PHP, C#, Dart
Frameworks & cloud
FastAPI, Flutter, React, Next.js, Vue, Nuxt 3/4, Laravel, Hono, SQLAlchemy, Drizzle ORM, Zod, PWA, Cloudflare Workers/D1/R2/Queues/Vectorize/AI Gateway, Docker
Data, security & tools
PostgreSQL, MySQL, SQL Server, Redis, RBAC, biometric authorization, audit logging, Tesseract.js, Vitest, GitHub Actions, Git, Jupyter, Pandas, NumPy, Supervisely

07 / Honors

Honors, awards & certificates

Machine Learning with Python

IBM · Coursera

PHP and Laravel, Batch 2

OSTAD

08 / Leadership

Leadership & community

Founder & Instructor

CodeSage

Founded an independent educational platform and taught programming and technology courses to over 100 university students, with technical mentorship and practical coding guidance.

Finance Head

Charity For Happiness

Managed financial activities and supported charitable work: financial assistance, food and winter clothing distribution, and direct on-field support for underprivileged communities.

Patrol Leader (Mate) · 2018

6th National COMDECA · Bangladesh Scouts

Led a Scout patrol at the National Community Development Camp in Haimchar, Chandpur, coordinating team logistics, disaster preparedness drills, and community service.

Sixer (Team Leader) · 2007

Manikganj District Cub Camporee · Bangladesh Scouts

Led a six-member Cub Scout unit representing Singair Pilot Primary School, coordinating camp logistics, team drills, and community service.

10 / Contact

Get in touch

I am looking for graduate research opportunities in autonomous agents, multimodal learning, and trustworthy AI. I would be glad to hear from faculty, labs, and scholarship committees.

References

Prof. Dr. Afroza NaharProfessor, Department of Computer Science, AIUB
Asst. Prof. Nuzhat TabassumAssistant Professor, Department of Computer Science, AIUB · Undergraduate thesis supervisor

Contact details for referees are in the CV.