Mitali Thapa
ML Engineer
Python
C++
Java
JavaScript
SQL
Node JS
C
Model Architecture & Training: 5/10
Data Pipeline & Feature Engineering: 5/10
Research Depth & Innovation: 5/10
Active 30 days ago
Invite to interview
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
A geospatial ML practitioner at a Senior level specializing in remote-sensing change-detection models and end-to-end mapping pipelines. The strongest proven skill is designing and integrating custom deep-learning change-detection architectures and dataset transforms, demonstrated by opencd models and transforms (for example open-cd/opencd/models/change_detectors/mtkd.py and open-cd/opencd/datasets/transforms/transforms.py). Public artifacts do not show robust production MLOps such as containerized CI/CD, comprehensive unit tests or secure secret management in automated deployment pipelines.
Phone
Technical skills
Languages
7
Python
C++
Java
JavaScript
SQL
Node JS
C
AI/ML
14
Scikit-learn
NumPy
OpenCV
Pandas
Albumentations
Streamlit
CUDA
Jupyter Notebook
TensorFlow
PyTorch
Keras
Stable Diffusion
Computer Vision
Reinforcement Learning
DevOps
8
AWS
GCP
Amazon S3
Kubernetes
Amazon EC2
GitHub
Rest API
Git
Analytics
3
Matplotlib
Power BI
Seaborn
Other
11
Express
FastAPI
MySQL
PyTorch C++
TensorFlow C++
Avalanche
Sensor Fusion
Recommender Systems
SLI/SLO/SLA
CNN
Image Segmentation
Timeline
AI Research Intern
•
Junior
GalaxEye
•
Internship
Researched and benchmarked computer vision and deep learning models for electro-optical satellite imagery. Performed EO image-processing experiments and analyzed failure cases to support model/architecture selection. Used Python tooling with deep learning frameworks and version control for research runs.
Python
PyTorch
Git
Academic Intern (Data Science)
•
Junior
Newton School
•
Internship
Analyzed AI-generated mock interview data together with human evaluator feedback to find performance patterns. Produced weekly analytical reports with recommendations to improve the AI assessment process. Supported mentoring activities covering SQL, Power BI, and Excel for students.
Python
SQL
Power BI
AI/ML Intern
•
Junior
Defence Research and Development Organisation (DRDO)
•
Internship
Built machine learning segmentation pipelines to process remote sensing imagery for glacial lake detection and hazard analysis. Conducted model evaluation and accuracy assessment to improve the reliability of glacial lake inventory outputs. Worked with Python-based ML workflows and remote sensing/geospatial data processing.
Python
TensorFlow
AI/ML Intern
•
Junior
Redblox.io
•
Internship
Developed an AI-based poster generation system using computer vision and diffusion-based techniques for automated layout creation. Designed backend data pipelines to place logos, captions, and visual elements dynamically. Implemented the service using a Python API and diffusion model tooling.
Python
Stable Diffusion
FastAPI
Sir Padampat Singhania University (SPSU)
Bachelor's Degree •
Computer Science Engineering (AI/ML)
Senior AI/ML Engineer
Confidence: High Research
Remote sensing and ML research engineer (senior-level) specializing in computer-vision change-detection model implementations and geospatial data pipelines. The strongest proven skill is building and integrating custom change-detection architectures and distillation workflows, demonstrated by opencd/models/change_detectors/mtkd.py and multiple custom decode/backbone modules. Public artifacts do not show production MLOps maturity, systematic experiment tracking, or unit / integration tests for reproducibility.
Model Architecture & Training
5/10
How well models are designed and trained
Custom model architectures and training-related logic are clearly implemented (multiple backbones, decoder heads, and a multi-teacher distillation framework), showing solid model engineering and paper reproduction effort but limited formal experiment management or extensive training automation.
Evidence
GalaxEye_Change_Detection/open-cd/opencd/models/change_detectors/mtkd.py: DistillSiamEncoderDecoder multi-teacher distillation implementation
GalaxEye_Change_Detection/open-cd/opencd/models/backbones/fcsn.py: FC_EF, FC_Siam_diff, FC_Siam_conc backbone implementations
GalaxEye_Change_Detection/open-cd/opencd/models/decode_heads/ban_utils.py: Transformer and MixFFN components used in decode heads
Data Pipeline & Feature Engineering
5/10
How data is prepared for models
A robust data pipeline for heterogeneous EO/SAR data is implemented with numerous custom transforms and loaders; notebooks show geospatial preprocessing and feature engineering for DEM and spectral indices.
Evidence
GalaxEye_Change_Detection/open-cd/opencd/datasets/transforms/transforms.py: large collection of MultiImg* transforms (resize, crop, augmentations)
GalaxEye_Change_Detection/open-cd/opencd/datasets/transforms/hetero_loading.py: LoadHeteroImagesFromFile and BinarizeLabels custom loaders
GLOF-Automated-Mapping-IHR/notebooks/Raster_To_ML.ipynb: load_xy_with_dem, flatten_xy and DEM feature stack logic
Experimentation & Evaluation
4/10
How results are measured and tested
Reasonable evaluation and experiment artifacts exist (custom metric class and evaluation notebooks), but there is limited evidence of formal experiment tracking, reproducible run configs across multiple runs, or automated ablation studies.
Evidence
GalaxEye_Change_Detection/open-cd/opencd/evaluation/metrics/scd_metric.py: SCDMetric implementation and compute_metrics
GalaxEye_Change_Detection/notebooks/05_phase5_training.ipynb: training and evaluation orchestration, log parsing and metric reporting
GLOF-Automated-Mapping-IHR/notebooks/Downsampling_Model.ipynb: RF evaluation, confusion matrix and feature importance
MLOps & Deployment
3/10
How models are shipped to production
Some serving/UX tooling and an inferencer class exist (Streamlit app for RF and an OpenCD inferencer), but no production-grade MLOps (CI/CD, model versioning, scalable serving, or drift monitoring) is evident.
Evidence
GLOF-Automated-Mapping-IHR/app/PrototypeUI.ipynb and app.py (Streamlit): UI/serving for Random Forest predictions
GalaxEye_Change_Detection/open-cd/opencd/apis/opencd_inferencer.py: OpenCDInferencer class for model inference
Computational Efficiency
2/10
How efficiently computing resources are used
Minor efficiency considerations appear (GPU checks, reduced iteration configs, some checkpointing and use of efficient attention blocks), but there is no systematic profiling, batching/throughput analysis, quantization, or distributed training evidence.
Evidence
GalaxEye_Change_Detection/notebooks/05_phase5_training.ipynb: GPU availability checks and config edits to limit iterations and batch sizes
GalaxEye_Change_Detection/open-cd/opencd/models/backbones/tinynet.py: imports and use of torch.utils.checkpoint hint at memory/speed trade-offs
Research Depth & Innovation
5/10
Depth of research and new ideas
The code reproduces and composes research-level ideas (multi-teacher distillation, transformer-based decode heads, cross-attention modules) showing authentic paper-implementation skill, though formal reproducibility evidence (detailed ablations, peer-reviewed results) is limited.
Evidence
GalaxEye_Change_Detection/open-cd/opencd/models/change_detectors/mtkd.py: explicit MTKD design invoking multiple teacher models
GalaxEye_Change_Detection/open-cd/opencd/models/decode_heads/bit_head.py and ban_utils.py: transformer encoder/decoder and custom attention/FFN components
Expertise
Computer Vision & Image Analysis• Senior
Geospatial AI & Satellite Imagery• Senior
AI Infrastructure & Optimization• Middle
Industries
Data & Analytics• Middle
Natural Resources• Senior
Science & Engineering• Middle
Technologies
SQL• since 2026
C++
MySQL
GCP
Stable Diffusion• since 2025
Reinforcement Learning• since 2026
Computer Vision• since 2026
Albumentations
TensorFlow• since 2025
Keras
Git• since 2026
PyTorch• since 2026
AWS• since 2026
Kubernetes
TensorFlow C++
PyTorch C++
Amazon EC2
SLI/SLO/SLA
CNN• since 2026
Image Segmentation• since 2026
CUDA
GitHub• since 2026
Amazon S3
Recommender Systems
Recommendations
- Extend and productionize geospatial change-detection models (deploy inference pipelines and optimized model serving for satellite/DEM inputs).
- Lead research/engineering efforts to implement and benchmark model distillation and lightweight deployable students for EO imagery (build on the MTKD and BITHead artifacts).
- Develop robust experiment tracking and reproducible training pipelines (W&B/MLflow integration, fixed run configs, CI for training/inference).
- Prototype scalable evaluation and monitoring for deployed models (latency/cost budgets, drift detection and automated re-evaluation).
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Senior Data Scientist
Confidence: Medium ML Practitioner
A geospatial ML practitioner at a Senior level specializing in remote-sensing change-detection models and end-to-end mapping pipelines. The strongest proven skill is designing and integrating custom deep-learning change-detection architectures and dataset transforms, demonstrated by opencd models and transforms (for example open-cd/opencd/models/change_detectors/mtkd.py and open-cd/opencd/datasets/transforms/transforms.py). Public artifacts do not show robust production MLOps such as containerized CI/CD, comprehensive unit tests or secure secret management in automated deployment pipelines.
Statistical Rigor
4/10
Correct use of statistics
Basic to moderate statistical practices are present - stratified splits, class imbalance experiments and standard classification reports - but there is little formal uncertainty quantification, no hypothesis tests with assumptions documented, and limited treatment of multiple-comparison or calibration.
Evidence
GLOF-Automated-Mapping-IHR/notebooks/Hybrid_Sampling_+_SMOTE.ipynb
GLOF-Automated-Mapping-IHR/notebooks/Downsampling_Model.ipynb
Data Wrangling & Cleaning
7/10
Preparing and cleaning data
Strong practical data-wrangling and custom loader work for geospatial data is evident - careful raster resampling, DEM feature extraction, custom dataset loaders and transforms to prevent runtime errors and label binarization.
Evidence
GalaxEye_Change_Detection/open-cd/opencd/datasets/transforms/hetero_loading.py
GalaxEye_Change_Detection/open-cd/opencd/datasets/basecddataset.py
GLOF-Automated-Mapping-IHR/notebooks/Raster_To_ML.ipynb
Exploratory Analysis & Visualization
6/10
Exploring and visualizing data
Exploratory analysis and visualization are pragmatic and question-driven with feature importance plots, confusion matrices and per-AOI evaluation; notebooks include interpretation and trade-off discussion rather than purely decorative plotting.
Evidence
GLOF-Automated-Mapping-IHR/notebooks/Downsampling_Model.ipynb
GLOF-Automated-Mapping-IHR/notebooks/Hybrid_Sampling_+_SMOTE.ipynb
GalaxEye_Change_Detection/open-cd/opencd/engine/hooks/visualization_hook.py
Predictive Modeling
6/10
Building models that predict
Predictive work spans classical ML (Random Forest with stratified splits, class-weighting and SMOTE experiments) and advanced deep-learning (custom change-detection architectures, distillation, custom heads and training hooks). Cross-validation and calibration are limited in notebooks, but the model engineering depth is substantial.
Evidence
GalaxEye_Change_Detection/open-cd/opencd/models/change_detectors/mtkd.py
GalaxEye_Change_Detection/open-cd/opencd/models/backbones/fcsn.py
GLOF-Automated-Mapping-IHR/notebooks/Downsampling_Model.ipynb
Business Insight & Impact
5/10
Turning analysis into business value
Analysis is tied to concrete domain goals - e.g., detection trade-offs for glacial-lake mapping and SLA versus cost in the autoscaling simulator - but formal estimation of business impact, cost-of-error quantification or stakeholder-level KPIs are partial rather than complete.
Evidence
GLOF-Automated-Mapping-IHR/README.md
Cloud-Autoscaling-RL/dashboard/app.py
GLOF-Automated-Mapping-IHR/notebooks/Downsampling_Model.ipynb
Reproducibility & Notebook Hygiene
3/10
Clean, repeatable analysis
Reproducibility is uneven: notebooks use seeds and save models, but there are many hard-coded Google Drive paths, imperative Colab environment patches, and ad-hoc in-notebook system modifications instead of pinned, containerized or CI-driven reproducibility practices.
Evidence
GalaxEye_Change_Detection/notebooks/05_phase5_training.ipynb
GLOF-Automated-Mapping-IHR/notebooks/Raster_To_ML.ipynb
Expertise
Data Science• Senior
Industries
Science & Engineering• Middle
Software• Middle
Technologies
Python• since 2025 • Senior
OpenCV• since 2026
Jupyter Notebook
Scikit-learn
Pandas• since 2026
NumPy• since 2026
Streamlit• since 2026
Recommendations
- Lead development of remote-sensing change-detection models and custom dataset pipelines, including dataset loaders, augmentations and evaluation metrics for EO imagery.
- Prototype and productionize model training pipelines with reproducible MLOps - containerization, pinned environments, CI and safe handling of secrets and credentials.
- Develop RL-driven infrastructure simulators and dashboards for capacity planning and SLA-cost trade-off analysis, leveraging the Cloud-Autoscaling-RL simulation and dashboard artifacts.
- Implement robust model evaluation practices: cross-validation, uncertainty quantification, calibration and automated performance monitoring for deployed geospatial models.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Senior AI / LLM Engineering (Agents)
Confidence: Medium Generalist
Remote-sensing computer vision engineer at a senior level with a strong focus on model architecture for change detection. The strongest proven skill is designing and implementing custom segmentation and transformer-based model components, as demonstrated by open-cd/opencd/models/decode_heads/ban_utils.py and open-cd/opencd/models/change_detectors/mtkd.py. Public code shows limited evidence of production hardening such as CI, unit tests, secure deployment practices or data governance.
Architecture & Organization
6/10
How the project is structured
Modular architecture and clear data flow for remote-sensing change detection pipelines; use of registry patterns and well-organized model/transform folders shows deliberate structure and plugin-style design.
Evidence
open-cd/opencd/datasets/transforms/transforms.py: TRANSFORMS.register_module() classes and modular pipeline transforms
open-cd/opencd/models/change_detectors/mtkd.py: multi-teacher distillation class registering complex dataflow between teacher/student models
open-cd/opencd/models/backbones/fcsn.py: custom backbone implementations following MODELS registry pattern
Code Quality & Standards
5/10
How clean the code is
Generally clean, idiomatic Python with type hints, asserts and docstrings; consistent naming and layered modules but limited evidence of automated tests or strict typing/enforcement across the codebase.
Evidence
open-cd/opencd/datasets/transforms/transforms.py: typed transform signatures, docstrings and assertions
open-cd/opencd/models/change_detectors/mtkd.py: use of typing (Tensor, ConfigType) and structured loss functions
open-cd/opencd/models/backbones/fcsn.py: clear docstring and consistent module-level organization
Engineering Culture
4/10
Teamwork and good practices
Consumes common ML engineering conventions (notebooks, configs, registry usage) and documents transforms and models, but repository lacks visible CONTRIBUTING/CI/tests and has limited collaboration metadata.
Evidence
notebooks/05_phase5_training.ipynb: detailed runnable notebook demonstrating experiment orchestration
open-cd/opencd/__init__.py & open-cd/opencd/apis/opencd_inferencer.py: consistent registry and API surface for inference
Technical Complexity
6/10
How hard the problems solved are
Implements non-trivial algorithms including custom attention/transformer blocks, transformer-based decoder heads and multi-teacher knowledge distillation; demonstrates moderate algorithmic sophistication targeted at performance and representational design.
Evidence
open-cd/opencd/models/decode_heads/ban_utils.py: CrossMultiheadAttention, MixFFN and BAN_BITHead with transformer components
open-cd/opencd/models/change_detectors/mtkd.py: Multi-Teacher Knowledge Distillation logic choosing teachers by change-area ratio
Security & Risks
2/10
Protection and risk handling
Little evidence of security-hardening or safe deployment patterns; several risky runtime patches and filesystem modifications occur inside notebooks which are dangerous when run without review.
Evidence
notebooks/05_phase5_training.ipynb: patches to site-packages, writing files into system packages and automated pip install commands
notebooks/05_phase5_training.ipynb: dynamic file writes to opencd/datasets/transforms/hetero_loading.py and edits to configs
Reliability & Fault Tolerance
3/10
Stability when things fail
Some defensive coding in transforms and try/except in visualization hooks exists, but there is limited systematic error handling, testing, CI or monitoring scaffolding for production reliability.
Evidence
open-cd/opencd/datasets/transforms/transforms.py: use of assertions and cache_randomness for deterministic transform behavior
notebooks/05_phase5_training.ipynb: try/except blocks around environment checks and patch application
Industries
Artificial Intelligence• Senior
Recommendations
- Develop advanced remote-sensing change-detection models and research prototypes that integrate custom backbones, transformer heads and distillation strategies.
- Build dataset preprocessing and augmentation pipelines for heterogeneous EO/SAR data using the existing transforms (transforms.py and hetero_loading) to create robust training inputs.
- Prototype model compression and lightweight inference flows for deployment, leveraging the existing distillation code as a starting point.
- Harden experiments for production by adding unit tests, CI, formal configuration management and removing risky runtime patching in notebooks.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: