Geetanshu Gupta
ML Engineer
Python
C++
SQL
JavaScript
C#
C
Model Architecture & Training: 4/10
Active 6 hours ago
Download CVCV
Overview
Technical skills
Projects
Timeline
Roles
Overview
A hands-on ML practitioner (middle level) focused on applied deep learning and computer vision prototypes with a strength in building end-to-end demo applications integrating detection models and lightweight language models.
Phone
Technical skills
Languages
6
Python
C++
SQL
JavaScript
C#
C
AI/ML
17
TensorFlow
MediaPipe
Transformers
YOLO
OpenCV
NumPy
Scikit-learn
Pandas
TF-Keras
PyTorch
LangChain
LangGraph
Streamlit
Hugging Face
EasyOCR
NLP
Model Context Protocol
DevOps
5
Azure
GitHub
Git
Rest API
Docker
Other
17
Matplotlib
FastAPI
MySQL
TensorFlow C++
Tkinter
Azure Cosmos DB
Postman
AI Agents
Deep Learning
Machine Learning
cert-manager
RAG
CNN
Embeddings
WebSockets
Computer Vision
GAN
Projects
Real-Time Agentic AI Voice System
Architected a model-agnostic real-time voice platform that integrates multiple AI providers via modular components. Implemented streaming voice communication with voice activity detection and barge-in handling for low-latency conversations. Added MCP-based tool orchestration so voice conversations can trigger backend operations and finance workflows.
AI-Powered Recruitment Intelligence Platform
Built a recruitment and resume analysis platform that parses resumes and computes an ATS score by comparing candidate content with job descriptions. Integrated generative AI and NLP techniques to rewrite weak resume bullets and recommend missing keywords for better alignment. Added a career recommendation component that matches skills and interests to roles and identifies skill gaps with learning resources.
Gesture Bridge
Designed and trained an AI-driven sign language translation system for high-throughput gesture recognition. Used a CNN-based deep learning approach on a large dataset to achieve real-time inference performance. Leveraged MediaPipe for gesture processing to improve accessibility outcomes for hearing-impaired users.
Timeline
AI/ML Engineer
·
Middle
PropVivo
·
Full-Time
Built agentic AI workflows using LLMs with LangGraph for multi-step task execution and tool orchestration. Implemented Model Context Protocol (MCP) tools and backend services so AI agents could invoke operations and finance functions via structured tool execution. Developed a real-time agentic AI voice platform using FastAPI and WebSockets with speech model integration, and deployed containerized services on Azure Container Apps. Added reliability and security features including automated failover, event-driven notifications, JWT authentication, RBAC, and IP-based rate limiting.
Python
LangGraph
Model Context Protocol
FastAPI
WebSockets
Azure
Docker
AI/ML Engineer (Intern)
·
Junior
PropVivo
·
Internship
Developed a RAG-based policy assistant using LLMs with LangChain and vector embeddings to support scalable backend APIs. Focused on reducing manual operations query handling by improving retrieval and response generation quality. Worked on integration of the assistant logic with service endpoints for production-style usage.
Python
RAG
LangChain
Embeddings
Middle AI/ML Engineer
Confidence: Medium ML Engineer
Applied computer vision and deep learning practitioner (Middle) specializing in prototyping real-time CV applications and simple generative-model training. The strongest proven skill is building end-to-end CV prototypes using YOLOv8 with OpenCV and glue logic for real-time gesture-to-text conversion as implemented in completeproject/button.py. There is little evidence of production-grade MLOps, systematic experiment tracking, robust evaluation suites, or scalable serving infrastructure.
Model Architecture & Training
4/10
How well models are designed and trained
Custom model code and training loops are present (GAN generator/discriminator, tf.keras DNN), and a YOLOv8-based real-time CV prototype is implemented. Architectures and loss definitions are standard and there is little evidence of advanced architecture design, ablation studies, or careful hyperparameter management.
Evidence
deep-learning-/Generating Artwork Using GANs PART 2.ipynb: build_generator, build_discriminator, discriminator_loss, generator_loss, train_step
Gesture-Bridge/completeproject/button.py: YOLO model loading and inference loop with smoothing/segmentation logic
deep-learning-/California_housing_prices (1) PART A.ipynb: tf.keras Sequential model definition and model.compile/train
Data Pipeline & Feature Engineering
2/10
How data is prepared for models
Basic data preprocessing and dataset creation are present (dataset batching and scaling) but there is no evidence of rigorous dataset versioning, augmentation strategies beyond simple normalization, nor pipeline orchestration for production datasets.
Evidence
deep-learning-/Generating Artwork Using GANs PART 2.ipynb: preprocess_data creates a tf.data.Dataset and batching
deep-learning-/California_housing_prices (1) PART A.ipynb: StandardScaler usage and train_test_split
Experimentation & Evaluation
2/10
How results are measured and tested
There are simple evaluation prints, validation runs and visualization functions (GAN image snapshots, printed metrics), but no experiment tracking, reproducible run configs, or systematic evaluation suites and no A/B or held-out experiment infrastructure.
Evidence
Gesture-Bridge/completeproject/evaluate_yolo.py: calls model.val() and prints precision/recall/mAP metrics
deep-learning-/Generating Artwork Using GANs PART 2.ipynb: generate_and_save_images called each epoch
deep-learning-/California_housing_prices (1) PART A.ipynb: training history and final MSE/R2 printing
MLOps & Deployment
1/10
How models are shipped to production
There is a simple tkinter launcher to run local scripts and evaluation scripts, but no production serving, model versioning, serialization strategy, CI/CD, or monitoring/drift detection.
Evidence
Gesture-Bridge/completeproject/launcher.py: tkinter UI that runs local python scripts via subprocess
Gesture-Bridge/completeproject/evaluate_yolo.py: local evaluation script expecting runs/detect/val folders
Computational Efficiency
1/10
How efficiently computing resources are used
Some pragmatic efficiency choices are present for prototypes (reduced webcam resolution, forcing CPU execution, and use of tf.function for training step) but there is no profiling, batching/throughput analysis, GPU/memory optimization, quantization, or other systematic efficiency work.
Evidence
Gesture-Bridge/completeproject/button.py: cap.set resolution reductions and model.overrides['device'] = 'cpu'
deep-learning-/Generating Artwork Using GANs PART 2.ipynb: @tf.function on train_step to improve training performance
Research Depth & Innovation
1/10
Depth of research and new ideas
Work shows applied implementations of standard techniques (DCGAN-like generator/discriminator, YOLOv8 usage) but no novel research contributions, no custom layers or paper-reproduction fidelity, and no published ablation or SOTA claims with reproducible evidence.
Evidence
deep-learning-/Generating Artwork Using GANs PART 2.ipynb: standard generator/discriminator and basic adversarial losses
Gesture-Bridge/completeproject/button.py: applied YOLOv8 inference loop for gesture-to-text conversion
Expertise
Computer Vision & Image Analysis· Middle
Generative Images & Diffusion· Middle
LLM· Junior
Industries
Health Care· Middle
Technologies
SQL
C++
MySQL
LangGraph· since 2026
LangChain· since 2026
Model Context Protocol· since 2026
cert-manager
WebSockets· since 2026
Embeddings· since 2026
Computer Vision· since 2026
AI Agents
NLP· since 2025
Azure· since 2026
TensorFlow· since 2025
Pandas
Git
PyTorch
Docker· since 2026
RAG· since 2026
MediaPipe· since 2025
TensorFlow C++
Streamlit
EasyOCR
Azure Cosmos DB
CNN· since 2025
Hugging Face
GitHub
Machine Learning· since 2026
Edge AI· mentioned only
Llama· mentioned only
MediaPipe· mentioned only
Python· mentioned only
TensorFlow· mentioned only
Recommendations
- Develop prototype-to-production workflows: add model serialization, inference API (FastAPI/Flask), and basic CI/CD for the CV apps to make demos reproducible and deployable.
- Adopt experiment tracking and reproducibility: add W&B or MLflow logging, fixed run configs, seed handling, and saved checkpoints for training experiments.
- Improve dataset and evaluation rigor: implement dataset versioning, augmentation pipelines, proper train/val/test splits, and scripted evaluation producing numeric reports and confusion matrices.
- Harden code for portability and security: remove hard-coded absolute paths, add CLI args/config files, dependency pinning, and defensive error handling for I/O and device selection.
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 hands-on ML practitioner (middle level) focused on applied deep learning and computer vision prototypes with a strength in building end-to-end demo applications integrating detection models and lightweight language models.
Statistical Rigor
2/10
Correct use of statistics
Minimal statistical rigor: basic metrics are computed but there are no assumption checks, uncertainty quantification, multiple-comparison controls, or causal analysis.
Evidence
deep-learning-/California_housing_prices (1) PART A.ipynb: computes MSE and R² but no residual analysis or assumption checks
deep-learning-/Generating Artwork Using GANs PART 2.ipynb: GAN training loop without uncertainty quantification or statistical evaluation
Data Wrangling & Cleaning
3/10
Preparing and cleaning data
Basic data wrangling is present (scaling, train/test split, simple preprocessing) but there is limited evidence of robust cleaning, provenance tracking, or leakage prevention.
Evidence
deep-learning-/California_housing_prices (1) PART A.ipynb: uses StandardScaler and train_test_split for preprocessing
Gesture-Bridge/completeproject/button.py: frame resizing and simple confidence-thresholding for detections
Exploratory Analysis & Visualization
3/10
Exploring and visualizing data
Exploratory visuals and some interpretation are present (prediction vs actual, training/validation loss, generated image grids) but analysis is mostly descriptive and lacks written, structured data storytelling.
Evidence
deep-learning-/California_housing_prices (1) PART A.ipynb: scatter plot of predicted vs actual and training/validation loss plots
deep-learning-/Generating Artwork Using GANs PART 2.ipynb: generate_and_save_images visualizations of GAN outputs
Predictive Modeling
3/10
Building models that predict
Predictive modeling skills are evident: Keras models, validation splits, and a custom GAN training loop with tf.function are used, but robust model evaluation (cross-validation, calibration, error analysis) and productionization patterns are limited.
Evidence
deep-learning-/Generating Artwork Using GANs PART 2.ipynb: custom generator/discriminator and training loop with tf.function
deep-learning-/California_housing_prices (1) PART A.ipynb: Keras regression model trained with validation_split and evaluation metrics
Business Insight & Impact
1/10
Turning analysis into business value
Little to no evidence of linking analysis to concrete business metrics, cost-of-error reasoning, or actionable recommendations in the code; product/readme statements exist but are not backed by domain-specific instrumentation or analysis in code.
Evidence
Ai-Powered-career-Recommendation/README.md: product feature descriptions and business-oriented claims (marketing text)
Reproducibility & Notebook Hygiene
2/10
Clean, repeatable analysis
Some notebook metadata and a requirements.txt are present, but reproducibility and engineering hygiene are weak: hardcoded absolute paths, no environment pinning, limited seed control, and no CI/pipeline scripts.
Evidence
Gesture-Bridge/completeproject/requirements.txt: lists dependencies but without exact pins for full reproducibility
deep-learning-/Generating Artwork Using GANs PART 2.ipynb: NOTEBOOK META present but no environment lockfile or DVC/pipeline scripts
Expertise
Data Science· Middle
Industries
Artificial Intelligence· Middle
Health Care· Middle
Technologies
Deep Learning
OpenCV
YOLO
Scikit-learn
Matplotlib
Transformers
NumPy
TF-Keras
GAN
Edge AI· mentioned only
Llama· mentioned only
MediaPipe· mentioned only
Python· mentioned only
TensorFlow· mentioned only
Recommendations
- Develop real-time computer vision prototypes and accessibility tools that integrate detection models (YOLO) with simple UX layers for demos and user testing.
- Build and iterate on generative model experiments (GANs) and small-scale model training pipelines, improving evaluation and reproducibility.
- Harden prototype code for production-readiness by removing hardcoded paths, adding configuration, device checks, and thorough error handling.
- Package notebooks into repeatable training scripts and CI-friendly pipelines with environment pinning and testable components.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Junior Backend Developer
Confidence: Medium Generalist
A generalist Python developer at a junior level who focuses on applied computer vision prototypes and interactive desktop demos. The strongest proven skill is building a real-time gesture-to-text pipeline combining YOLOv8 detection and a lightweight causal language model as implemented in completeproject/button.py. Public code does not demonstrate backend systems, APIs, databases, testing, or production-grade observability and deployment practices.
API Design
How well APIs are designed
Not evidenced in public code
Data Layer & Database
Working with databases
Not evidenced in public code
Scalability & Performance
2/10
Handling load and speed
Minimal runtime and frame-level performance considerations exist (resolution reduction, detection thresholds, frame-stability logic) but no production-grade scaling, connection pooling, or load-handling strategies.
Evidence
Gesture-Bridge/completeproject/button.py: cap.set(cv2.CAP_PROP_FRAME_WIDTH/HEIGHT) and LETTER_CONF_THRESHOLD / STABLE_LETTER_FRAMES frame-stability logic
Gesture-Bridge/completeproject/button.py: model.overrides['device'] = 'cpu' and single-process frame loop
System Architecture
2/10
Overall system structure
Simple script decomposition with a small Tkinter launcher to run components; there is minimal service decomposition, no config/secret management, and no inter-service contracts.
Evidence
Gesture-Bridge/completeproject/launcher.py: init_main_window and run_script launching other scripts via subprocess
Gesture-Bridge/completeproject/button.py and evaluate_yolo.py: separate scripts for detection and evaluation
Security & Auth
1/10
Protecting data and access
Very light security posture: basic try/except and file-path checks only; hard-coded paths and subprocess invocation present security and secrets risks.
Evidence
Gesture-Bridge/completeproject/button.py: hard-coded yolo_model_path and try/except around model load
Gesture-Bridge/completeproject/launcher.py: subprocess.run(["python", script_path]) with absolute paths and no sanitization
Reliability & Observability
2/10
Stability and monitoring
Basic reliability practices such as try/except and resource cleanup are present, but there is no structured logging, timeouts/retries with backoff, graceful shutdown orchestration, or observability/metrics.
Evidence
Gesture-Bridge/completeproject/button.py: try/except on model load and cap.release() / cv2.destroyAllWindows() cleanup
Gesture-Bridge/completeproject/evaluate_yolo.py: try/except around evaluation and printed metric output
Expertise
Backend AI & LLM· Junior
Python· Junior
Industries
Artificial Intelligence· Middle
Health Care· Junior
Technologies
Python· since 2025 · Junior
Rest API
FastAPI· since 2026
Tkinter
Rich· mentioned only
Text· mentioned only
Recommendations
- Develop end-to-end proof-of-concept CV prototypes and accessibility demos that integrate detection models with simple UIs (desktop or kiosk apps).
- Build small production wrappers for models (HTTP microservice) to gain experience with APIs, request handling, and secure model serving.
- Improve operational hygiene: add structured logging, graceful shutdown, retries with backoff, and basic metrics to make prototypes production-ready.
- Add automated tests and a minimal CI pipeline; replace hard-coded paths and subprocess usage with configurable, sanitized execution patterns
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: