Devmallya Karar
AI Engineer
7+ years ML exp
Python
SQL
Model Architecture & Training: 4/10
Data Pipeline & Feature Engineering: 4/10
Experimentation & Evaluation: 4/10
Active 8 days ago
+91 (9535) 032238 Invite to interview
Message
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
Applied ML engineer at a middle level specializing in fraud and anomaly-detection pipelines with a strength in end-to-end model development for imbalanced datasets. The strongest proven skill is hands-on imbalance handling and model evaluation, demonstrated by multiple notebooks that implement oversampling/undersampling (SMOTE/ADASYN/RandomOverSampler/RandomUnderSampler), systematic GridSearchCV experiments and threshold/ROC analysis. What is not evidenced is production-grade engineering - there is little modular production code, no CI/experiment tracking, serving or monitoring infrastructure in the public artifacts.
Technical skills
Python• Middle
SQL
Python
Django
Pydantic
FastAPI
AI/ML
Claude Code
Cline
CNN
CrewAI
Cursor
Embeddings
Hallucination
Keras
NLP
NumPy
Pandas
Prompt Engineering
Scikit-learn
Transformers
TF-Keras
Torchvision
Statsmodels
YOLO
LightGBM
Jupyter Notebook
Deep Learning
Computer Vision• 7y+
ResNet• 7y+
TensorFlow• 7y+
Time Series Forecasting• 7y+
BERT• 5y+
NER• 5y+
Vertex AI• 5y+
XGBoost• 5y+
Fine-tuning• 4y+
LLM• 4y+
Reinforcement Learning• 4y+
AWS Bedrock
Claude
Few-Shot Learning
LangChain
LangGraph
Mistral
PyTorch
RAG
Langfuse
LangSmith
Model Context Protocol
DevOps
Azure
Docker
Heroku
Vector
Rest API• 8y+
CI/CD• 7y+
AWS• 5y+
GCP• 4y+
Analytics
Matplotlib
Seaborn
Cybersecurity
GDPR
Robotics
PID Control
QA
Postman• 8y+
Timeline
Lead AI Engineer
•
Lead
Bayer AG
•
Full-Time
Owned the end-to-end architecture of Bayer’s enterprise machine-translation platform, including domain-aware routing and speech-to-text/text-to-speech capabilities. Designed an agentic conversational system that combines retrieval and multi-agent collaboration with real-time streaming. Built and orchestrated multiple production agents on AWS using LangGraph/LangChain, including secure tool-discovery via MCP and OAuth-based access control, with human approval checkpoints for sensitive actions. Implemented LLM evaluation and observability with LangSmith and Langfuse based on live tool-call traces, and delivered a camera translation feature using OCR and LLMs.
AWS
LangGraph
LangChain
RAG
Model Context Protocol
LangSmith
Langfuse
LLM
Lead AI Consultant
•
Lead
HCLTech
•
Full-Time
Delivered a first-generation domain-adaptive machine-translation system for Bayer using LangChain/LangGraph agents, RAG, and routing across multiple engines. Implemented an AI-generated-text detection and blocking layer using Amazon Bedrock (Claude) with few-shot prompting, reducing non-compliant content by the stated amount. Built a PII-safe medical document parser on AWS using Bedrock and FastAPI, applying RAG chunking for field extraction under healthcare data-protection constraints. Developed email intent classification with BERT-based models and AWS services, and built response generation workflows using Bedrock models with PyTorch and LangChain.
AWS Bedrock
AWS
LangGraphsince 2024
LangChainsince 2024
RAGsince 2024
BERT
FastAPI
Few-Shot Learning
PyTorch
LLM
Claude
Mistral
AI/ML Engineer - Lead Assistant Manager
•
Lead
EXL
•
Full-Time
Led a team of ML engineers and data scientists delivering AI, computer-vision, and NLP products across multiple insurance and financial-services clients. Built document-oriented GenAI solutions for summarization and key-value extraction, taking multiple solutions from proof of concept to production. Fine-tuned LLM-based extractors and improved extraction quality using BERT-enhanced approaches, leveraging GCP for model workflows. Implemented a Mask R-CNN pipeline for extracting information from embedded charts and mentored junior data scientists.
Fine-tuning
LLMsince 2022
BERT
GCP
Computer Vision
Liverpool John Moores University
Master's Degree •
Machine Learning and Artificial Intelligence
Machine Learning Engineer
•
Middle
Quantiphi
•
Full-Time
Shipped a DistilBERT-based named-entity recognition model for medical entities extracted from insurance-claim forms. Developed ensemble classifiers over clinical text entities using XGBoost to improve classification quality versus a single-model baseline. Built BERT-NER proofs of concept for document extraction using AWS Textract and document-layout approaches. Delivered a real-time medical-invoice table detection model on GCP Vertex AI with low per-page latency.
BERTsince 2021
NER
XGBoost
AWSsince 2021
Vertex AI
Computer Vision
Data Scientist - Associate Product Development
•
Middle
HARMAN International
•
Full-Time
Designed multivariate deep-learning forecasting models in TensorFlow to improve sales forecast accuracy versus a statistical baseline. Built a 3D hand-gesture recognition system for an automotive OEM using a ResNet architecture and deployed it on-device with TensorFlow Lite on GCP. Created reproducible CI/CD pipelines for scalable model deployment, reducing release cycle time. Performed analysis on high-dimensional socio-economic data using GIS spatial-classification techniques to support product targeting decisions.
TensorFlow
ResNet
CI/CD
Time Series Forecasting
Computer Visionsince 2019
Vertex AI
Software Engineer
•
Middle
Hewlett-Packard
•
Full-Time
Developed and deployed a collaborative-filtering recommendation system for HP’s e-commerce platform to improve click-through behavior. Built a customer-support chatbot using Rasa with Python to automate routine queries and reduce tier-1 support load. Deployed ML models to production using Django along with GCP Vertex AI to serve live traffic on the e-commerce site. Validated API performance and reliability with Postman across many model endpoints before release.
Vertex AIsince 2018
Postman
Rest API
Middle AI/ML Engineer
Confidence: Medium ML Engineer
Applied ML engineer at a middle level specializing in fraud and anomaly-detection pipelines with a strength in end-to-end model development for imbalanced datasets. The strongest proven skill is hands-on imbalance handling and model evaluation, demonstrated by multiple notebooks that implement oversampling/undersampling (SMOTE/ADASYN/RandomOverSampler/RandomUnderSampler), systematic GridSearchCV experiments and threshold/ROC analysis. What is not evidenced is production-grade engineering - there is little modular production code, no CI/experiment tracking, serving or monitoring infrastructure in the public artifacts.
Model Architecture & Training
4/10
How well models are designed and trained
Practical model-building across classical ML, gradient-boosted trees and neural nets; shows end-to-end training, hyperparameter search and a custom RL network and training loop but no novel research-level architectures or extensive optimizer/schedule engineering.
Evidence
Credit_Card_Fraud_Detection_DL_1.ipynb: autoencoder model definition (input_layer -> encoder -> decoder -> autoencoder) with callbacks and training loop
Credit_card_fraud_Devmallya_Karar (2).ipynb: extensive GridSearchCV hyperparameter tuning for RandomForest/XGBoost/KNN across sampling strategies
Credit_Card_Fraud_Detection_via_DQNs.ipynb: class DQN and Agent.learn custom training loop with soft_update and replay buffer
Data Pipeline & Feature Engineering
4/10
How data is prepared for models
Solid applied data pipeline and imbalance handling work: scaling, power-transform, many sampling strategies (oversampling/undersampling/SMOTE/ADASYN) and feature analysis; pipelines are notebook-driven rather than production ETL.
Evidence
Credit_card_fraud_Devmallya_Karar (2).ipynb: PowerTransformer usage (PowerTransformer(method='yeo-johnson')) and creation of X_train_transform/X_test_transform
Credit_card_fraud_Devmallya_Karar (2).ipynb: imbalanced-learn sampling pipelines (RandomOverSampler, SMOTE, ADASYN, RandomUnderSampler) applied before model training
Time Series forecasting to predict the volumes Final.ipynb: stationarity tests and SARIMAX model selection functions (adfuller_test, best_sarima_model)
Experimentation & Evaluation
4/10
How results are measured and tested
Reasonable experimentation and evaluation practices: stratified CV, GridSearchCV, custom thresholding/ROC analysis, confusion matrices and SHAP for feature importance; missing formal experiment tracking and reproducible config files.
Evidence
Credit_card_fraud_Devmallya_Karar (2).ipynb: repeated use of GridSearchCV with StratifiedKFold across models and sampling variants (e.g., randomforest_model_2, xgboost_model_1)
Credit_card_fraud_Devmallya_Karar (2).ipynb: optimal_cutoff and draw_roc functions used to select thresholds and plot ROC/metrics
Credit_Card_Fraud_Detection_DL_1.ipynb: shap.TreeExplainer and shap.summary_plot to inspect model feature importances
MLOps & Deployment
2/10
How models are shipped to production
Minimal MLOps/serving evidence: model checkpointing and saving are used, and there are notes about Kaggle/Colab commands; lacks deployment/service code, CI/CD, monitoring or model versioning.
Evidence
Credit_Card_Fraud_Detection_DL_1.ipynb: tf.keras.callbacks.ModelCheckpoint(filepath='best_model.h5') and saving/loading best_model.h5
Credit_Card_Fraud_Detection_DL_1.ipynb: use of kaggle CLI and Colab setup commands in notebook for data download
Computational Efficiency
2/10
How efficiently computing resources are used
Some awareness of computational choices (GPU-enabled XGBoost, torch.device selection) but no measured efficiency work, profiling, batching studies or quantization.
Evidence
Credit_card_fraud_Devmallya_Karar (2).ipynb: XGBClassifier(..., tree_method='gpu_hist') indicating GPU usage for training
Credit_Card_Fraud_Detection_via_DQNs.ipynb: device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') and .to(device) usage for networks/tensors
Research Depth & Innovation
2/10
Depth of research and new ideas
Applied implementations of known methods (autoencoder anomaly detection, DQN agent) but no original algorithmic contributions or research-grade reproductions with ablations and published comparisons.
Evidence
Credit_Card_Fraud_Detection_DL_1.ipynb: autoencoder-based anomaly detection pipeline with threshold selection and evaluation
Credit_Card_Fraud_Detection_via_DQNs.ipynb: DQN class, ReplayBuffer and Agent.learn custom RL loop
Expertise
AI Agents & Agentic Workflows• Middle
MLOps & Model Lifecycle• Junior
Technologies
Python• Middle
SQL
Cursor
LangGraph
Rest API• 8y+
LangChain
Claude
GCP• 4y+
Claude Code
Model Context Protocol
XGBoost• 5y+
Vertex AI• 5y+
FastAPI
Fine-tuning• 4y+
Embeddings
Reinforcement Learning• 4y+
Prompt Engineering
Computer Vision• 7y+
NLP
NER• 5y+
Cline
Langfuse
LangSmith
Heroku
Azure
Mistral
AWS Bedrock
CI/CD• 7y+
Transformers
TensorFlow• 7y+
Django
PyTorch
AWS• 5y+
Docker
CrewAI
LLM• 4y+
RAG
BERT• 5y+
ResNet• 7y+
Pydantic
Torchvision
Vector
Few-Shot Learning
Hallucination
CNN
Time Series Forecasting• 7y+
Recommendations
- Refactor notebooks into modular Python packages and reusable modules (data preprocessing, model training, evaluation) with unit tests and clear entry points for productionization.
- Add experiment tracking and reproducibility (W&B/MLflow runs, fixed config files, seed control) and a CI pipeline to validate models and training scripts.
- Implement a lightweight serving prototype (FastAPI or KServe/BentoML) and basic monitoring (metrics, drift checks) for the best-performing model to validate production constraints.
- Harden the RL work by adding stable training loops, evaluation episodes, deterministic seeds and clearer reward/episode logging; consider leveraging stable-baselines3 or Ray RLlib for scale and reproducibility.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Data Scientist
Confidence: Medium ML Practitioner
A middle-level ML practitioner focused on applied predictive modeling and time-series forecasting with strong hands-on experience. The strongest proven skill is end-to-end model development and tuning for time-series and tabular problems, demonstrated by SARIMA-based forecasting pipelines and large-scale classification work with hyperparameter search and tree-ensemble models. Not evidenced are production engineering practices such as modularized reusable code, environment pinning, CI/CD, automated tests, and data/version control.
Statistical Rigor
5/10
Correct use of statistics
Uses appropriate statistical tests and discusses assumptions (ADF, Shapiro, decomposition and normality) but uncertainty quantification is limited and there are occasional metric/method misuse issues.
Evidence
Time-Series-forecasting-to-predict-the-volumes/Time Series forecasting to predict the volumes Final.ipynb: test_stationarity / adfuller_test / shapiro_normality_test
Classification-Problem---Insurance/Classification Problem - Insurance Devmallya Karar (2).ipynb: Central Limit Theorem and confidence interval calculations
Data Wrangling & Cleaning
6/10
Preparing and cleaning data
Solid practical data wrangling: missing-value handling, outlier removal, encoding and scaling are implemented; some operations are notebook-bound with hard-coded local paths and potential reproducibility/leakage risks.
Evidence
Classification-Problem---Insurance/Classification Problem - Insurance Devmallya Karar (2).ipynb: outlier removal (IQR) and get_dummies / StandardScaler pipelines
Time-Series-forecasting-to-predict-the-volumes/Time Series forecasting to predict the volumes Final.ipynb: month-name to numeric conversion and datetime index setup
Exploratory Analysis & Visualization
6/10
Exploring and visualizing data
Exploratory analysis is thorough and accompanied by written interpretation (seasonality, ACF/PACF, distributions, class counts); visualizations are used to motivate modeling choices.
Evidence
Time-Series-forecasting-to-predict-the-volumes/Time Series forecasting to predict the volumes Final.ipynb: plot_data_properties, seasonal_decompose and year-by-year monthly plots with commentary
Classification-Problem---Insurance/Classification Problem - Insurance Devmallya Karar (2).ipynb: extensive countplots, distribution plots and grouped analyses with text explanations
Predictive Modeling
5/10
Building models that predict
Good breadth of predictive modeling (SARIMA/SARIMAX, ARIMA, LSTM, logistic, tree ensembles, XGBoost, LightGBM) and hyperparameter search (RandomizedSearchCV, hyperopt), but evaluation hygiene and production-readiness (robust error analysis, calibration, clear CV pipelines) are uneven.
Evidence
Time-Series-forecasting-to-predict-the-volumes/Time Series forecasting to predict the volumes Final.ipynb: best_sarima_model grid-search function and SARIMAX forecasting pipeline
Classification-Problem---Insurance/Classification Problem - Insurance Devmallya Karar (2).ipynb: hyperopt objective, RandomizedSearchCV for RandomForest, XGBoost and LightGBM training and model saving
Stock-price-Prediction-using-LSTM/Stock price Prediction using LSTM.ipynb: LSTM model definition, scaling, training and prediction pipeline
Business Insight & Impact
3/10
Turning analysis into business value
Some problem framing and discussion of precision-recall tradeoffs for the insurance use-case, but limited explicit business impact quantification, cost-of-error analysis or deployment/operational considerations.
Evidence
Classification-Problem---Insurance/Classification Problem - Insurance Devmallya Karar (2).ipynb: Problem Statement framing for an insurance client and discussion of precision-recall tradeoffs in model selection
Time-Series-forecasting-to-predict-the-volumes/Time Series forecasting to predict the volumes Final.ipynb: explanation of MAPE and selection rationale
Reproducibility & Notebook Hygiene
2/10
Clean, repeatable analysis
Work is notebook-first and reproducible for exploration but lacks production hygiene: hard-coded local paths, no environment pin files, minimal modularization, and only ad-hoc model pickling.
Evidence
Multiple notebooks (e.g. Time Series forecasting to predict the volumes Final.ipynb, Classification Problem - Insurance Devmallya Karar (2).ipynb) use absolute local file paths like 'E:/PIP_Devmallya/...' and save artifacts to local paths
Classification-Problem---Insurance/Classification Problem - Insurance Devmallya Karar (2).ipynb: model saved via pickle but no requirements/conda/Docker or data versioning present; presence of random_state in train_test_split shows partial reproducibility effort
Industries
Commerce• Middle
Financial Services• Middle
Technologies
Deep Learning
Jupyter Notebook
YOLO
Scikit-learn
Seaborn
Matplotlib
LightGBM
Pandas
NumPy
Keras
TF-Keras
Statsmodels
Recommendations
- Develop prototype production pipelines: refactor notebook code into modular Python packages, add unit tests, and adopt CI to make models deployable.
- Add reproducibility artifacts: provide requirements/conda files, container (Docker) images and remove hard-coded local paths so colleagues can re-run and deploy.
- Focus on MLOps and monitoring: build model serving, scheduling (Airflow/Dagster) and post-deployment monitoring for forecasts and classifiers.
- Improve evaluation rigor: add proper probability calibration, class-imbalance handling, robust cross-validation schemes and clearer business-cost/error analyses.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: