News
Submitted Beyond the Update Boundary: Explanation Reproducibility in Clinical Time-Series Models to the AI4TS Workshop @ AAAI 2027.
Towards Reliable ML Evaluations: A Temporal-Validity Audit for Trustworthy Model Updates accepted at MIDI 2026 (Springer LNCS).
Research
My research is on deep learning for time series and sequential data, and on keeping these models trustworthy when the data change: how sequence models such as LSTMs, GRUs, temporal convolutional networks and attention-based hybrids learn from data that unfolds over time, how to make their predictions more accurate, how to tell real shift in the data from artifacts of how it was recorded, and whether their explanations stay reliable after an update. The same questions arise wherever models learn from streams of data: in wearable health monitoring, brain–computer interfaces (EEG), human–computer interaction, sensor monitoring in power and nuclear systems, and language-model systems, where a model's explanations and outputs need to be checked against independent evidence that may itself go out of date. More broadly, I welcome related problems in machine learning, time series and trustworthy AI.
I have worked on hybrid ensembles for forecasting and load prediction, and on sequence models trained on fifteen years of hospital time series (MIMIC-IV): auditing the data before a model update is evaluated (MIDI 2026), and measuring how reproducible a model's explanations are after it is updated (under review). In industry, I also build production AI systems: real-time voice agents, retrieval-augmented LLM applications and computer-vision pipelines.
Publications and preprints
MIDI 2026Towards Reliable ML Evaluations: A Temporal-Validity Audit for Trustworthy Model Updates
A temporal-validity audit that checks, before a model update is evaluated, whether apparent shift in MIMIC-IV ICU time series (2008–2022) reflects real change or the database recording process. Re-scoring the same models on audited data reversed one update comparison from +0.051 to −0.030 AUROC. Springer LNCS. Accepted, to appear.
STI 2024Hybrid Deep Learning Approach for Load Profile Prediction
BiLSTM, BiGRU and TCN layers with self-attention, combined by an XGBoost regressor, for real-time load forecasting in smart grids. 6th International Conference on Sustainable Technologies for Industry 5.0, Dhaka.
ICISET 2024Enhancing Prediction Accuracy using Ensemble Deep Learning Methods on Time Series Data
An ensemble of BiLSTM, BiGRU and TCN with an XGBoost meta-learner for general time series forecasting. International Conference on Innovations in Science, Engineering and Technology (IEEE).
PreprintAudit Before You Update: Recording Artifacts and the Limits of Bounded Model Updates in MIMIC-IV ICU Prediction
Sequence models (LSTM and gradient-boosted trees) on ICU time series from 2008–2022: audit recording artifacts before any update, detect drift without labels, and compare bounded against full model updates with pre-specified non-inferiority margins.
Under reviewBeyond the Update Boundary: Explanation Reproducibility in Clinical Time-Series Models
Measures how reproducible a time-series model's explanations are after it is updated: a two-stream model (LSTM over hourly ICU physiology, MLP over treatment context) updated across three eras with its encoder frozen or fine-tuned, at four learning rates and five seeds. Under review at the AI4TS Workshop @ AAAI 2027.
Live demos
Kotha AI ● Live
A Bangla voice receptionist for dental clinics that books, reschedules and cancels real appointments over a live call, with a staff portal showing every transcript.
- Speech recognition chosen by measurement: 48/50 Bangladeshi names correct vs 38/50 for the alternative
- Double-booking made impossible in the database, not the prompt
- Automatic fallback for every provider, with a watchdog for silent failures
AI Messenger ● Live
A family messenger with an AI-twin mode: when someone is busy, their twin replies in their own tone, using retrieval over their journals and past conversations and a Big Five personality profile.
- Retrieval-augmented memory with FAISS and per-user privacy boundaries
- Relationship-aware replies (parent, sibling, partner)
- Real-time chat over WebSockets
Selected projects
Physician ID from prescription photos
Identifies the prescribing doctor from a phone photo of a prescription page using the printed letterhead: ResNet-50 embeddings, OCR features and nearest-neighbour retrieval. 93.1% leave-one-out accuracy on the images that can be matched.
Video to recipe with vision-language models
Turns a cooking video into a structured recipe: selects key frames, captions them with BLIP, transcribes the audio with Whisper, and has an LLM combine both into a recipe with dietary variants.
Matritya: maternal health app
An Android app for expecting mothers in Bangladesh with trimester-based diet and exercise plans, symptom-based disease prediction, vaccination reminders and a community space.
Genome variant classification
Classifies ClinVar single-nucleotide variants as benign, pathogenic or of uncertain significance with a 1D CNN over allele encodings and genomic position.
Experience and education
Independent research, with my undergraduate co-authors
Data auditing and model updating under temporal shift, leading to the MIDI 2026 paper and the AI4TS submission.
AI Developer, BDCallingIT
Production ReAct-based RAG pipelines, computer-vision APIs and voice agents.
Freelance AI Engineer, Outlier AI
Reference solutions and adversarial prompts for LLM code generation; ranking model outputs for technical accuracy.
BSc in Computer Science and Engineering
Patuakhali Science and Technology University, Bangladesh
Undergraduate Researcher, Patuakhali Science and Technology University
Hybrid deep learning models for time series forecasting, leading to the ICISET 2024 and STI 2024 papers.
Ankur Undergraduate Scholarship, Ankur International
Champion, Barisal Divisional Hackathon (Hackverse)
9th place, Mujib Borsho IT Carnival
Tools
Contact
I'm happy to talk about research collaborations, graduate study and AI engineering. Email is the quickest way to reach me.