HomeResearch projects

Work on realresearch problems.

49 research topics in five tracks. Each starts with a simple question. Open the brief when you want the technical detail.

Browse projects How to join

How to join

Three steps to a project team

Projects are part of the Research Methods Acceleration program. A supervisor approves every team.

01
Join the program
Join the waitlist, then enrol in Research Methods Acceleration when the batch opens.
02
Finish the foundations
Complete the basics and a short readiness check, then request a topic.
03
Join a team
The supervisor reviews your request and places you in a small team.

What you produce

A supervised research package: a reproducible baseline, evaluation results and a written report.

Each week

About 6–8 hours of your own work, plus team check-ins with your supervisor, across the 12 weeks.

Timeline

Your supervisor shares the milestone plan when the team is formed.

Enrolling does not guarantee a topic, a place on a team, publication or authorship. Before work starts, the supervisor checks data permissions, skills and a working baseline.

Five tracks

Research areas

Open projects accept requests now. Proposed topics are planned and open later, once a supervisor is confirmed.

Reliable language AI
Scientific question answering, citation checks, retrieval and evidence control across Bangla and English.
14 projects
Climate and geospatial
Flood, drought and heat mapping, tested across different events, regions and uncertain reference data.
12 projects
Industrial vision
Product inspection and defect detection that keeps working when lighting, devices or internet connections change.
12 projects
Reliable systems and analytics
Broken sensors, federated learning, changing data and safety checks for AI tools and code.
10 projects
Health data
Public health questions answered with open data and reproducible methods.
1 project

All projects

Find a question you care about

Reliable language AI

Open

Can Bangla AI Know When It Is Wrong?

Test whether an AI system can tell when its answer is likely wrong, even after students switch between Bangla, English or technical terms.

Research question: Does confidence calibrated on one language remain useful when students change language, script or technical terminology?
What the team tests: A risk-controlled answer/abstain policy evaluated after matched, expert-checked language transformations.
Data: BEnQA and P6Jiggasha (licence-dependent) plus a held-out transformed subset with two annotators.
Compared against: Base-model answering, simple confidence thresholding, post-hoc calibration and selective prediction methods.
How results are measured: Risk-coverage curves, selective accuracy, calibration error, coverage gaps by language, confidence intervals resampled at seed-question level.

Team of 4 students

Reliable language AI

Open

Build a Research Assistant Whose Citations Can Be Checked

Check whether the sources an AI assistant cites really support what it says, when the question and the papers are in different languages.

Research question: When query and evidence use different languages, which claims acquire plausible citations that do not actually support them?
What the team tests: Separate unsupported, partially supported and fully supported statements, including quantities and experimental conditions.
Data: A rights-cleared corpus in one technical field; initially 400-800 claim-passage pairs as a planning range, with expert annotation distinguishing relevance from entailment.
Compared against: Citation-generation baseline, retrieval-only relevance checking, suitable NLI-based verification and a strong language-model judge; measure judge agreement against humans.
How results are measured: Claim-support macro-F1, unsupported-citation rate, citation completeness, human agreement and cost per checked answer.

Team of 4 students

Reliable language AI

Open

Can AI Understand the Bangla-English Students Actually Use?

Find out whether mixing Bangla and English in a question makes AI pick the wrong sources, reason badly, or both.

Research question: Does code-switching impair evidence retrieval, reasoning over correct evidence, or both?
What the team tests: A controlled intervention matrix (original vs mixed-language query, correct vs retrieved context, native vs translated technical terms).
Data: Expert-reviewed variants of permitted scientific QA examples plus a small consented set of naturally phrased student queries; artificially mixed text must not be described as authentic student language.
Compared against: Multilingual retriever, translation-first retrieval, hybrid lexical/vector retrieval and an applicable normalization method.
How results are measured: Retrieval recall at k, answer accuracy with gold vs retrieved context, paired performance change, latency and native-speaker naturalness ratings.

Team of 4 students

Reliable language AI

Open

When AI Should Say "The Documents Do Not Tell Us"

Teach an AI assistant to admit when the documents it found do not contain the answer.

Research question: Can an assistant distinguish genuinely sufficient evidence from relevant but incomplete documents after the query language changes?
What the team tests: Evidence removal at the level of necessary facts, not just empty-context prompts; test robustness across language and unseen document sources.
Data: A bounded scientific corpus with answerable questions and independently validated incomplete-context counterparts, including realistic retrieval errors.
Compared against: Always answer, prompt-only abstention, confidence thresholding and relevant published evidence-sufficiency approaches.
How results are measured: Unsupported answer rate at matched coverage, refusal precision/recall, performance on answerable cases, and cost.

Team of 4 students

Climate and geospatial

Open

Flood Maps That Work Beyond the Flood They Were Trained On

Build AI flood maps that still work on a new flood, using only a small number of hand-labelled images.

Research question: With a fixed small annotation budget, can adaptation improve performance on a new flood event without leaking target-event information?
What the team tests: Compare event-aware adaptation against strong conventional and pretrained baselines, separating spatial from cross-event transfer.
Data: NASA flood dataset coverage and reference labels; verify available events, geography and labels before promising Bangladesh-wide coverage.
Compared against: Traditional backscatter thresholding, supervised segmentation baseline, appropriate pretrained model and credible adaptation baseline; never tune on final-event test labels.
How results are measured: Flood-class IoU/F1, false positives on permanent water, event-wise performance, annotation cost and event-level uncertainty.

Team of 4 students

Climate and geospatial

Open

Flood Monitoring When Satellite Observations Are Missing

Find which gaps in satellite data break flood detection, and test a method that copes with them.

Research question: Which kinds of missing observations cause the largest detection failures, and can a gap-aware method improve reliability?
What the team tests: Evaluate real acquisition gaps and sustained outages, not only randomly masked pixels; radar and optical observations have different failure modes.
Data: A prevalidated temporal satellite stack with independent flood labels; the NASA extent dataset is a starting point, not proof that the required dense time series exists.
Compared against: Available-sensor-only model, last-observation baseline, simple imputation and missingness-aware fusion across several held-out events.
How results are measured: Accuracy by gap length/type, uncertainty, performance recovery after observations resume, and computational cost.

Team of 4 students

Industrial vision

Open

Finding Product Defects Despite Changing Factory Lighting

Cut false alarms in AI product inspection when factory lighting changes, without missing tiny defects.

Research question: Can a small allowed recalibration set reduce false alarms under lighting changes while preserving sensitivity to tiny defects?
What the team tests: A predeclared calibration-budget protocol and a transferable calibration strategy, tested on unseen categories/conditions; the lighting problem itself is explicitly present in MVTec AD 2.
Data: MVTec resources subject to licence and official evaluation rules; do not redistribute restricted labels or use the test server as a validation oracle.
Compared against: A strong patch-feature detector, a lightweight alternative and applicable test-time/calibration approaches.
How results are measured: Image and pixel measures, defect-level recall at fixed false-alarm budgets, calibration-set size, confidence intervals and category-wise failure analysis.

Team of 4 students

Industrial vision

Open

Can a Small Device Inspect Products Reliably?

Check whether shrinking an AI model to run on a small device makes it miss small or rare defects.

Research question: Does model compression disproportionately remove detection of small or rare defects, and can calibration recover reliability at a fixed device budget?
What the team tests: Actual device measurements and risk/latency tradeoffs, rather than parameter counts alone.
Compared against: Original model, standard post-training quantization, an applicable quantization-aware or distillation method, and an uncompressed smaller architecture.
How results are measured: Defect-level recall, false alarms, calibration, end-to-end latency, peak memory and measured energy only if a credible measurement setup exists.

Team of 4 students

Health data

Open

Secondary Data Analysis in Public Health

Analyse open public health data using methods that others can check and repeat.

Team of 4 students

Reliable language AI

Proposed topic