Anuj Kumar
ML Engineer
Python
TypeScript
SQL
JavaScript
Model Architecture & Training: 4/10
Active 6 days ago
Download CVCV
Overview
Technical skills
Projects
Timeline
Roles
Overview
Frontend UI engineer (middle level) focused on polished component work and visual design tokens. The strongest proven skill is building UI primitives and local state patterns, exemplified by the custom global toast system in src/hooks/use-toast.ts and a consistent tokenized theme in src/index.css and tailwind.config.ts. There is limited evidence of formal testing, large-scale architecture decisions, advanced server-state handling or explicit accessibility implementations in public code.
Phone
Technical skills
Languages
4
Python
TypeScript
SQL
JavaScript
AI/ML
19
Streamlit
Machine Learning
NumPy
Pandas
Scikit-learn
Keras
TF-Keras
Gemini
LangChain
OpenCV
Transformers
YOLO
TensorFlow
PyTorch
SciPy
OpenAI
Google Colab
MediaPipe
Hugging Face
DevOps
3
GitHub
Git
Azure
Frontend
9
React.js
Tailwind CSS
React Query
Zod
React Router
React Hook Form
Framer Motion
Recharts
Vite
Other
15
Pydantic
FAISS
Wikipedia
ESLint
AI Agents
Deep Learning
RAG
Embeddings
Fine-tuning
Prompt Engineering
Recommender Systems
Semantic Search
Computer Vision
NLP
Transfer Learning
Projects
Built a multi-agent fact verification system that extracts claims, retrieves relevant evidence from Wikipedia, ranks sources, and generates an evidence-grounded verdict with confidence. Implemented a retrieval pipeline using semantic chunking, vector search, and MMR-based retrieval to improve relevance and reduce hallucinations. Integrated LLM tooling and Gemini-based answering via LangChain and Transformers.
Designed and trained a Random Forest model to optimize traffic signal timing for a four-way intersection based on live traffic conditions. Implemented YOLO-based real-time vehicle detection to classify high- and low-traffic lanes and feed decisions to the signal timing logic. Delivered the workflow via a Streamlit interface and reported an R² score of 0.977.
Created a KNN-based book recommendation system as part of a supervised machine learning project. Built the pipeline to train and apply the KNN model for recommendations and evaluated the approach using standard ML practices from the project scope.
Timeline
GLA University
Bachelor's Degree ·
Computer Science Engineering (Hons.) with specialization in AI & Data Analytics
Middle AI/ML Engineer
Confidence: High ML Engineer
ML engineer (mid-level) focused on building end-to-end, project-level ML apps and demos - strongest at assembling pipelines and simple model training with UI frontends. The proven strength is building end-to-end prototypes and deployable demos, as shown by the Streamlit apps plus model training and serialization in Cat-and-Dog-Classification/imageDataGenerator.py and book-recommendation-system/.../train_model.py. There is limited evidence of production-grade MLOps, experiment tracking, custom research, rigorous evaluation or deployment automation in public code.
Model Architecture & Training
4/10
How well models are designed and trained
Standard model construction and training are present (CNN built with Keras/TF, KNN-based recommender). No custom layers or advanced training loops, but there is correct use of model.fit, early stopping and model serialization.
Evidence
Cat-and-Dog-Classification/imageDataGenerator.py: Sequential CNN definition, model.compile, model.fit and model.save
book-recommendation-system/Book recomendation using KNN/train_model.py: NearestNeighbors model.fit and joblib.dump
Data Pipeline & Feature Engineering
3/10
How data is prepared for models
Basic but correct data preparation and filtering (datetime parsing, deduplication, pivot tables, sparse matrix creation); pipelined end-to-end preprocessing for apps.
Evidence
vireo-audio-support-analytics/Main.py: ticket['created_at'] to_datetime, .dt.to_period('W') grouping and drop_duplicates
book-recommendation-system/Book recomendation using KNN/train_model.py: creation of pivot_table and filtering active users and popular books
Experimentation & Evaluation
2/10
How results are measured and tested
Minimal experimentation and evaluation artifacts; basic evaluation call and a reported validation accuracy but no structured experiment tracking, baselines, or reproducible experiment configs.
Evidence
Cat-and-Dog-Classification/imageDataGenerator.py: model.evaluate(val_data) and EarlyStopping callback usage
Cat-and-Dog-Classification/README.md: reported validation accuracy (88%) referenced but without formal eval pipeline or train/test split management in code
MLOps & Deployment
2/10
How models are shipped to production
Simple serving patterns using Streamlit and local model serialization; no production deployment, model versioning, monitoring, or CI/CD evidence.
Evidence
vireo-audio-support-analytics/app.py: Streamlit UI calling run_analysis() and presenting tables/metrics
book-recommendation-system/Book recomendation using KNN/app.py: Streamlit frontend using joblib.load('model.pkl')
Computational Efficiency
2/10
How efficiently computing resources are used
Some attention to efficiency at the application level (sparse matrices, generator-based data feeding) but no measured optimization, GPU/memory profiling or quantization.
Evidence
book-recommendation-system/Book recomendation using KNN/train_model.py: use of scipy.sparse.csr_matrix for the pivot table
Cat-and-Dog-Classification/imageDataGenerator.py: ImageDataGenerator usage with batch_size and on-the-fly augmentation
Research Depth & Innovation
1/10
Depth of research and new ideas
Little to no research-level work; standard models and pipelines without paper re-implementation, custom algorithms or novel methods.
Evidence
Cat-and-Dog-Classification/imageDataGenerator.py: standard Sequential Keras model (no custom layers or novel components)
book-recommendation-system/Book recomendation using KNN/train_model.py: classical KNN approach, no paper-implementation or ablation study
Expertise
LLM· Junior
Technologies
SQL
LangChain
OpenCV
FAISS
YOLO
Fine-tuning
Embeddings
Prompt Engineering
Computer Vision· since 2026
AI Agents
NLP· since 2026
Transfer Learning
Azure
Google Colab
Transformers
TensorFlow
Keras
Git· since 2024
PyTorch
Gemini
RAG
MediaPipe
Pydantic
Hugging Face
GitHub· since 2024
Semantic Search
Recommender Systems
Wikipedia
Machine Learning· since 2024
CNN· mentioned only
Python· mentioned only
Recommendations
- Develop interactive prototype features and dashboards (Streamlit) that integrate model inference with clear input validation and error handling.
- Work on experiment tracking and reproducibility (e.g., add MLflow or W&B, fixed random seeds and train/validation/test splits) to turn demos into reproducible experiments.
- Add automated tests and CI for data pipelines and model serving, plus remove hardcoded local paths to make projects portable and production-ready.
- Focus on incremental MLOps skills: model versioning, lightweight deployment (Docker), and basic monitoring/metrics for served models.
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
An ML practitioner at a middle level who builds compact end-to-end ML prototypes with Streamlit frontends and model training pipelines. The strongest proven skill is assembling working prototype ML systems that combine preprocessing, model training and UI delivery as shown in book-recommendation-system/Book recomendation using KNN/train_model.py and Cat-and-Dog-Classification/imageDataGenerator.py. Not evidenced are production-grade engineering practices such as automated testing, CI/CD, robust error handling, scalable data pipelines, or rigorous model-validation workflows.
Statistical Rigor
2/10
Correct use of statistics
Minimal statistical rigor; no hypothesis tests, uncertainty quantification, multiple-comparison awareness, or documented assumption checks.
Evidence
vireo-audio-support-analytics/Main.py: calculate_business_outcome returns point estimates without confidence intervals or statistical tests
book-recommendation-system/Book recomendation using KNN/train_model.py: model training and filtering performed with no cross-validation or baseline comparisons
Data Wrangling & Cleaning
4/10
Preparing and cleaning data
Solid small-scale data wrangling and feature preparation steps (deduplication, datetime parsing, filtering) but several brittlenesses exist and there are limited provenance or leakage-prevention controls.
Evidence
vireo-audio-support-analytics/Main.py: drop_duplicates, pd.to_datetime, weeks period extraction, groupby aggregations
book-recommendation-system/Book recomendation using KNN/train_model.py: filtering active users and popular books and pivot_table creation
Exploratory Analysis & Visualization
2/10
Exploring and visualizing data
Exploratory outputs are limited to Streamlit tables and basic UI; EDA lacks diagnostic visualizations and written interpretation of findings.
Evidence
vireo-audio-support-analytics/app.py: Streamlit displays of weekly_digest and weekly_agents as dataframes without analytic commentary
book-recommendation-system/Book recomendation using KNN/app.py: Streamlit UI returns recommendations but contains no exploratory plots or interpretation notes
Predictive Modeling
3/10
Building models that predict
Predictive modeling implementations exist (KNN recommender and CNN) and include sensible ML APIs (NearestNeighbors, model.fit, early stopping) but lack rigorous evaluation protocols, cross-validation, calibration, or error analysis.
Evidence
book-recommendation-system/Book recomendation using KNN/train_model.py: NearestNeighbors training and model persistence via joblib
Cat-and-Dog-Classification/imageDataGenerator.py: Sequential CNN model with early stopping, fit and model.save
Business Insight & Impact
4/10
Turning analysis into business value
Some business linkage is present - repeat-contact rate and cost are computed and surfaced in the UI - but the treatment of error costs, FP/FN tradeoffs and actionability is shallow.
Evidence
vireo-audio-support-analytics/Main.py: calculate_business_outcome computes repeat_rate and repeat_cost
vireo-audio-support-analytics/app.py: Streamlit metrics showing Total Contacts, Repeat Contact Rate, and Cost of Repeat Contacts
Reproducibility & Notebook Hygiene
1/10
Clean, repeatable analysis
Low reproducibility and notebook hygiene - hardcoded local paths, no pinned environment beyond a simple requirements file, no seeds, no data/versioning or CI evidence.
Evidence
book-recommendation-system/Book recomendation using KNN/train_model.py: hardcoded Windows paths for data files
Cat-and-Dog-Classification/imageDataGenerator.py: local model.save and absolute dataset paths
Expertise
Data Science· Middle
Technologies
Deep Learning
Python· since 2024 · Middle
Scikit-learn
SciPy
Pandas
NumPy
TF-Keras
Streamlit· since 2024
OpenAI
Analytics· mentioned only
CNN· mentioned only
Recommendations
- Use to build rapid end-to-end ML prototypes and internal dashboards that demonstrate model value to stakeholders.
- Improve model evaluation discipline by adding cross-validation, baseline comparisons, calibration, and systematic error analysis.
- Refactor projects for reproducibility: remove hardcoded paths, pin environments, add deterministic seeds, and add data/versioning or lightweight CI.
- Partner with engineers to productionize models: add unit tests, input validation, error handling, and deployable packaging (Docker, CI/CD)
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Frontend Developer
Confidence: Medium UI Engineer
Frontend UI engineer (middle level) focused on polished component work and visual design tokens. The strongest proven skill is building UI primitives and local state patterns, exemplified by the custom global toast system in src/hooks/use-toast.ts and a consistent tokenized theme in src/index.css and tailwind.config.ts. There is limited evidence of formal testing, large-scale architecture decisions, advanced server-state handling or explicit accessibility implementations in public code.
UI Component Architecture
6/10
How interface parts are built
Component boundaries and a small internal design system are present; there is evidence of custom hooks and intentional tokenization but no large-scale component library or storybook-level systemization.
Evidence
Responsive & Cross-browser
5/10
Works on all screens and browsers
Responsive practices and dark mode tokens are used with fluid typography in key files, but there is no explicit RTL/i18n or advanced feature-detection code.
Evidence
AnujKumar1204/anuj-kumar-portfolio/tailwind.config.ts: container padding, screens and darkMode configuration
AnujKumar1204/veritas-ai/Frontend_react/src/components/Hero.module.css: clamp() for responsive headings
AnujKumar1204/anuj-kumar-portfolio/src/index.css: dark theme token set and base responsive rules
Performance Optimization
3/10
Speed of the interface
Minor measured performance-aware patterns appear (lightweight CSS animations, a memory-based toast store to limit rerenders) but there is no evidence of profiling, bundle analysis, virtualization or advanced code-splitting strategy.
Evidence
AnujKumar1204/anuj-kumar-portfolio/src/hooks/use-toast.ts: memoryState + listeners pattern to manage toast updates without forcing global React renders
AnujKumar1204/anuj-kumar-portfolio/src/index.css: well-scoped animations and utilities to avoid heavy repaints
AnujKumar1204/anuj-kumar-portfolio/vite.config.ts: presence of Vite config shows tool awareness (basic build tooling)
Accessibility & Semantics
2/10
Usable for everyone
Visual and semantic CSS is thoughtful, with loading indicators and visible focusable controls implied, but there is little explicit a11y code such as ARIA attributes, keyboard management in custom widgets, or CI a11y tooling.
Evidence
AnujKumar1204/veritas-ai/Frontend_react/src/components/ResultCard.module.css: loading spinner and interactive header styling but no ARIA clues
AnujKumar1204/veritas-ai/Frontend_react/src/components/InputSection.module.css: interactive controls styled but no explicit accessibility attributes
AnujKumar1204/anuj-kumar-portfolio/eslint.config.js: ESLint configured but no a11y plugin present
State Management & Data Flow
4/10
Managing data in the app
There is deliberate local state management (a reducer-based toast system and typed Supabase types), showing server-state awareness at a typing level, but no advanced server-state strategies like request cancellation, optimistic updates with rollback or state machines.
UX & Visual Polish
5/10
Look and feel quality
High visual polish is evident with design tokens, animations, spinners and skeleton-like elements, but UX edge handling (comprehensive loading/error/empty states and undo flows) is not consistently implemented across artifacts.
Evidence
AnujKumar1204/anuj-kumar-portfolio/src/index.css: extensive design tokens, animation keyframes and utility classes
AnujKumar1204/veritas-ai/Frontend_react/src/components/InputSection.module.css: spinners, disabled states and interactive button feedback
AnujKumar1204/veritas-ai/Frontend_react/src/components/ResultCard.module.css: loading row, spinner and result state styles
Industries
Artificial Intelligence· Middle
Technologies
TypeScript· Middle
Tailwind CSS
React.js
Vite
React Query
ESLint
Zod
Framer Motion
Recharts
React Hook Form
React Router
Recommendations
- Own and extend a component library or design system work where tokenization and component composition are required.
- Build small-to-medium single page applications that integrate server-state flows using React Query or similar, emphasizing optimistic updates and cancellation handling.
- Add automated accessibility checks and keyboard/ARIA handling for custom widgets to raise accessibility quality.
- Introduce unit and integration tests around custom hooks and critical flows, and add basic bundle-analysis performance checks
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: