Open to Graduate Research & Collaboration
5 Preprints Under ReviewEmpirical AI for low-resource languages, multimodal reasoning & verifiable evaluation.
I'm Khan Raiyan Ibne Reza, undergraduate researcher at North South University — Research Assistant (1 yr) · Teaching Assistant (3 semesters). My research investigates where foundation models fail on low-resource Bengali, dialectal retrieval, and multimodal vision-language reasoning, paired with resilient full-stack systems engineering.
5
Research preprints
Under peer review
7
Shipped systems
Open-source codebases
1 Yr
Research Assistant
Dept. of ECE · NSU
3
Semesters TA
Dept. of History & Philosophy
North South University
BSc in Computer Science & Engineering · Dept. of ECE
Research Assistant · 1 yr
Teaching Assistant · History & Philosophy ×3 semesters (200+ students)
Top 10 Solvio AI · Top 20 InnovativeX
ICPC Asia Dhaka Regional '22, '23 · ECE Capstone 2nd Runner-Up
Academic & Systems Credentials
Verifiable record across research and engineering.
Every metric traces to a university transcript, benchmark repository, or public GitHub codebase.
5
Research Preprints
All 5 preprints under peer review / arXiv-indexed
7
Shipped Systems
Production codebases with public source code
117
Completed Credits
BSc in Computer Science & Engineering at NSU
200+
Students Mentored
Undergraduate Teaching Assistant ×3 semesters (DHP)
Spotlight
Featured work
Flagship benchmark · 85,979 instances
KrishokChat: provenance-traceable Bengali agricultural benchmark
01 · Language & Retrieval
Low-Resource Bengali NLP & RAG Diagnostics
Measuring semantic drift across 6 regional dialects, failure modes in dense vs. sparse retrieval pipelines, and verifiable factual grounding over domain-specific corpora.
Explore 3 preprints02 · Multimodal & Reasoning
Vision-Language & Document Understanding
Evaluating cross-modal consistency in geometric reasoning (diagram comprehension) and diagnosing multimodal models on historical land records with base-16 fractional arithmetic.
Explore 2 preprintsFlagship preprints
Two benchmarks that highlight our core evaluation methodology.
View all 5 preprintsarXiv:2606.29243 · Preprint Under Review
KrishokChat: Bengali Agricultural Benchmark & Grounding Diagnostics
85,979-instance multi-task evaluation suite from 284 government publications across 13 institutions and 6 regional dialects. Evaluates dialectal retrieval degradation, tabular reasoning, and domain-grounded advisory.
arXiv:2608.15223 · Preprint Under Review
TRACE-BN: Tutoring Dialogue Reasoning for Sub-1B Edge Models
4,099 curriculum-guided tutoring traces for pedagogical step verification and code-mixed Bangla-English error correction deployed locally on edge devices (378.3 MB GGUF).
All Five Research Preprints & Benchmarks
Where Does Retrieval Fail? Evaluating RAG Architectures for Agricultural Advisory
TRACE-BN: Transferring Bangla-English Tutoring Behavior to a Sub-1B Offline Language Model
ChitraMiti: A Benchmark for Cross-Modal Consistency in Bengali Geometric Reasoning
Production systems
Shipped, not slideware.
Seven verifiable production codebases — every card links to a live GitHub repository.
View all 7 systemsKrishokChat
Role · Lead AI & Systems Architect
Safety-aware Bengali agricultural advisory RAG with crop vision disease diagnostics.
Nirapod Probaho (AegisFlow)
Role · Full-Stack & Multi-Agent Lead
AI-orchestrated multi-agent disaster resilience and relief distribution platform.
PawCare
Role · Sole Developer
Veterinary telemedicine and multimodal clinical assistant application.
Let's turn a hard Bengali-NLP problem into a measured result.
Open to graduate research discussions, evaluation baselines, benchmark inquiries, and systems engineering collaboration. Replies within 48 hours.