Anamul Hoque Emtiaj

Department of CSE, BUET · Supervisor: Prof. Anindya Iqbal · Interdisciplinary AI systems

• Auditing bank statements in natural language [C1] — parse statement PDFs with a multi-tier parser (deterministic PyMuPDF for tabular pages, LlamaParse for the rest), repair parsing errors heuristically, standardise the result into PostgreSQL, and audit it in natural language with a SQL agent. We compared context-injection strategies for the agent — zero context, RAG context, and execution feedback with an agentic validation loop — and how far newer model generations reduce the need for a complex architecture.
• Building-code compliance checking (in progress) — a code-generation agent that produces reusable and one-time checkers for building-code clauses against building models. Running experiments toward a conference submission in the LLM4Code category.

UIUC++ Summer Research in Software Engineering, remote · Supervisors: Prof. Darko Marinov (UIUC), Prof. Wenxi Wang (University of Virginia)

• Working on formal verification of Rust programs using Verus: specification testing for LeetCode and Codeforces problems, and test-case generation.

Department of CSE, BUET · Supervisor: Prof. M. Sohel Rahman · Computational biology

• CRISPR-Cas9 off-target prediction [J1] — designed CRISMER to predict off-target sites from a DNA–guide-RNA sequence pair: a multi-branch CNN that extracts k-mer features from the encoded pair, coupled with transformer blocks that capture long-range dependencies across it. Applied DeepSHAP and Integrated Gradients to attribute predictions back to mismatch positions, and used the resulting model to drive guide-RNA optimization.
• Achieved a 2–8% improvement in PR-AUC and F1 over recent state-of-the-art baselines on genome-wide off-target datasets, after surveying CNN, RNN/LSTM/GRU, transformer, and BERT-based architectures. Supported by the RISE BUET Students' Research Grant.

Dhaka, Bangladesh · engaged with two North American startup clients