Researcher in Reliable Machine Learning

Tanveer Hossain Munim

Reliable Machine Learning · Uncertainty Quantification · Trustworthy Evaluation

I study how calibration and evaluation choices affect model comparisons, and develop uncertainty methods for deployment.

View Research Download CV

Research Profile

Research Focus & Background

My current research covers calibration-aware evaluation of rare-event forecasts, regional uncertainty for object counting, exact robustness certificates for forecast rankings, and conformal risk control that stays valid as an open-vocabulary segmentation model's vocabulary grows.

I am a Senior Research Engineer at AI GeoLAB, where I work on evaluated, human-in-the-loop systems for Bengali land-record digitisation. Previously, at RIMES, I led backend, data-pipeline, and ML architecture for national-scale meteorological and early-warning systems. I hold a B.Sc. in Computer Science and Engineering from BUET.

Research Interests

Reliable Machine Learning Uncertainty Quantification Calibration Trustworthy AI Evaluation Conformal Prediction
Tanveer Hossain Munim

Contact

Academic Foundation

Education

  • April 2018 – September 2023

    B.Sc. in Computer Science and Engineering

    Bangladesh University of Engineering and Technology (BUET)

    CGPA 3.62/4.00 · Undergraduate thesis: ShasthoSheba · Supervisor: Prof. A. B. M. Alim Al Islam · GRE 329 (Q170, V159) · IELTS 8.0

Research Output

Publications & Manuscripts

Work on reliable evaluation, uncertainty quantification, geospatial ML, and earlier human-centred computing.

Submitted and under review

  • 2026

    A Free Knob: Decoupling Calibration and Predictive Skill in Threshold-Based Evaluation

    Munim, M. T. H., Saiem, B. A., Sany, A.-A., & Hashem, T.

    Under review at ICLR 2027 · Datasets & Benchmarks

    A matched-calibration audit for thresholded rare-event benchmarks. On released CasCast checkpoints, the cascade-over-backbone CSI gap falls from 0.1601 to 0.0339 after monotone recalibration; across 450 SEVIR contrasts, 51 reverse sign.

    CalibrationPythonReliable ML Code ↗ arXiv ↗
  • 2026

    ShasthoSheba: Leveraging an mHealth Solution for Providing Healthcare Service to Orphanages

    Munim, M. T. H., Anan, M. T. T., Galib, F.-Z.-I., Saiem, B. A., & Islam, A. B. M. A. A.

    ACM Transactions on Computer-Human Interaction · Under review

    Human-Centred Computing

Peer-reviewed publication

  • 2024

    eDakterBari: A human-centered solution enabling online medical consultation and information dissemination for resource-constrained communities in Bangladesh

    Eliza, I. J., Urmi, M. A., Anan, M. T. T., Munim, M. T. H., Galib, F.-Z.-I., & Islam, A. B. M. A. A.

    Heliyon, 10(1), e23100

    Human-Centred Computing DOI ↗

Work in progress

  • 2026

    Anchored Risk Profiles: Exact Robustness Certificates for Fixed-Threshold Forecast Rankings under Acceptable Corrections

    Munim, M. T. H.

    Manuscript in preparation

    Exact certificates for whether forecast rankings survive every acceptable calibration-improving monotone correction.

    CalibrationPythonReliable ML
  • 2026

    Confidence-Gated Extraction of Handwritten Bengali Land Records with a Fine-Tuned Vision-Language Model

    Munim, M. T. H., Ahmed, R., Karim, M. A., Lee, K., & Wu, J.

    Industry-track manuscript in preparation

    Document AIQwen2.5-VLReliable ML
  • 2026

    Click Recentring: Regional Count Intervals from a Few Hand-Counted Tiles

    Munim, M. T. H., Ikram, F. Z., & Hassan, A. M.

    Manuscript in preparation

    A headroom audit shows that no signal computed from the image narrows count-scaled conformal intervals for counting. In click recentring, a person counts four random tiles per image; recentring every other region keeps the coverage guarantee and narrows dense-region intervals by a median of 15% under dataset shift and 18% on unseen FSC-147 categories.

    Computer VisionConformal PredictionReliable ML

Invited Talks & Presentations

  • 29–30 April 2026

    “Climate–Agriculture Risk Modeling: Sri Lanka”

    Climate Services User Forum, South Asian Hydromet Forum (SAHF) · Malé, Maldives

    On behalf of RIMES

  • 9 October 2025

    “Machine Learning for Extreme-Event Detection in Agrometeorology”

    Capacity Building Training, National Center of Meteorology · UAE

    On behalf of RIMES

Research in Practice

Selected Research & Applied Work

Selected empirical research, applied research, and research software.

FreeKnob Audit — Calibration vs Predictive Skill

A model-agnostic audit for threshold-based rare-event evaluation. Matched monotone calibration separates predictive skill from threshold placement without retraining or changing spatial order.

Calibration Python Reliable ML

Regional Uncertainty for Object Counting

A headroom audit of post-hoc count intervals, and click recentring: a few hand-counted tiles per image recentre every region's interval while keeping split-conformal coverage. Evaluated on three crowd datasets and FSC-147.

Computer Vision Conformal Prediction Reliable ML

Anchored Risk Profiles

An exact rational-arithmetic framework for testing whether a fixed-threshold forecast ranking survives every acceptable monotone recalibration. Manuscript in preparation.

Calibration Python Reliable ML

Vocabulary Shift in Open-Vocabulary Segmentation

Conformal risk control for open-vocabulary segmenters breaks when class names are added after calibration: on ADE20K, a threshold set for 10% risk at 150 names reaches up to 15.4% risk at 845 names in released CAT-Seg models. Scoring each label against a fixed reference bank of names, without retraining, keeps one calibration valid as the vocabulary grows, on three benchmarks and five segmenters. Supervised by Prof. Muhammad Abdullah Adnan, BUET; manuscript in preparation.

Computer Vision Conformal Prediction Reliable ML

Engineering

Selected Production Systems

Deployed infrastructure, applied ML systems, and open-source contributions.

Reliable Bengali Land-Record Digitisation

At AI GeoLAB, a fine-tuned Qwen2.5-VL-3B, deployed for government use, scores 95-97% on the record fields of villages unseen in training, and its own token probabilities decide which drafts need only a quick check. A separate segmentation and number-reading pipeline recovers and correctly numbers 91.7% of plots on 22 held-out cadastral sheets.

Document AIGeospatial MLPyTorchQwen2.5-VL

GCF CDIS — Timor-Leste National Meteorological Platform

Led backend, data-pipeline, and ML architecture for a national meteorological platform processing more than 1 TB/day from five NWP models and two satellites. The system also supports operational CAP v1.2 alerting.

Apache AirflowDjango / DRFDockerPostgreSQL / PostGISxarray / NetCDF

Research Toolkit

Methods & Technical Foundations

Reliability & Evaluation

  • Conformal prediction
  • Calibration and benchmark auditing
  • Robust model comparison
  • Pre-registration and reproducibility

Statistical Methods

  • Bootstrap inference
  • Weighted and Mondrian conformal methods
  • Bayesian modelling with PyMC
  • Exact rational-arithmetic analysis

Vision & Language Models

  • PyTorch and Hugging Face
  • VLM fine-tuning with QLoRA
  • Segmentation, OCR, and crowd counting
  • Hybrid retrieval and local inference

Geospatial & Production Systems

  • xarray, GDAL, Rasterio, and PostGIS
  • ERA5, IMERG, GloFAS, SEVIR, and WeatherBench2
  • Airflow, Celery, and Docker
  • Python, Rust, TypeScript, and SQL

Research & Practice

Research and Professional Experience

  • April 2026 – Present

    Senior Research Engineer

    AI GeoLAB Ltd

    Developing evaluated, human-in-the-loop methods for handwritten Bengali land records and cadastral map digitisation.

  • July 2024 – July 2026

    Senior Software Engineer

    Regional Integrated Multi-Hazard Early Warning System (RIMES)

    Led backend, data-pipeline, and ML architecture for Timor-Leste CDIS; designed CAP v1.2 alerting and the data unification behind Bangladesh FFWC's live 15-day flood forecasts.

  • June 2021 – August 2022; June 2023 – August 2024

    Backend Developer; later Tech Lead

    Interactive Cares

    Led an eight-person engineering team as the learning platform grew from 8,000 to 100,000 users.

Credentials & Recognition

Awards & Affiliations

NST Fellowship Bangladesh National Science and Technology Fellowship · 2026
NVIDIA Inception MIRA AI · 2025
Accelerating Asia Selection, team member · Top 9 of 500+ · 2023
Bangladesh Mathematical Olympiad Champion · IMO team shortlist · 2016

Let's Connect

Get in Touch

I welcome conversations about graduate research and collaborations in reliable machine learning.

The quickest way to reach me is directly by email. I usually respond within 24 hours.

Email Me

Also find me on

Open To

Graduate Research Opportunities Research Collaborations Reliable ML Climate & GeoAI Research