Madhumitha Kolkar
Madhumitha Kolkar
Seasoned Machine Learning Engineer with 3.3 years of professional experience working with a specialization in Natural Language Processing, Computer Vision, Deep Learning and Generate AI.
- Role
Machine Learning Engineer
- Years of Experience
3 years
- Professional Portfolio
View here
Skillsets
- Keras - 3 Years
- LangChain/Llama - 2 Years
- Cnn - 3 Years
- Dialogflow, Microsoft Bot Framework - 3 Years
- MediaPipe - 3 Years
- LangChain - 3 Years
- FastAPI - 3 Years
- flask - 3 Years
- Streamlit - 4 Years
- spaCy - 4 Years
- Nltk - 4 Years
- Python - 6 Years
- GCP - 3 Years
- Generative AI - 2 Years
- Large Language Models (LLMs) - 3 Years
- Computer Vision - 3 Years
- Deep Learning - 3 Years
- Natural Language Processing (NLP) - 3 Years
- Machine Learning - 3.3 Years
- Pytorch - 3 Years
- TensorFlow - 3 Years
Professional Summary
3Years
- Apr, 2021 - Feb, 20242 yr 10 months
Machine Learning Engineer
Mercedes Benz Research and Development, India - Dec, 2020 - Apr, 2021 4 months
Data Scientist
Deloitte
Applications & Tools Known
OpenCV
NumPy
Dialogflow
Mediapipe
Streamlit
Pandas
PyTorch
scikit-learn
Android SDK
MySQL
MongoDB
Git
Work History
3Years
Machine Learning Engineer
Mercedes Benz Research and Development, IndiaApr, 2021 - Feb, 20242 yr 10 months
- Pioneered NLP Chatbot (GCP): Championed NLP Chatbot for Mercedes-Benz service appointments, reducing unnecessary appointments by 15% & boosting profit by $2M annually, leading to a 15% increase in customer satisfaction.
- Enhanced User Intent Classification (LSTM): Integrated scaled-up LSTM network, significantly improving user intent classification with F1 scores exceeding 0.85.
- Introduced Automotive Lexicons: Strengthened the model by introducing domain-specific lexicons, reducing human intervention for user queries by 10%
- Championed Ensemble with BERT: Advocated for and co-developed a hybrid ensemble model using a fine-tuned BERT model and LSTM network for superior performance resulting in an increased F1 score by 12%.
- Data Augmentation & Feature Engineering: Pioneered innovative data augmentation (SMOTE) and domain-specific feature engineering, boosting model robustness and accuracy by 20%.
- Engineered accuracy and reduced misinterpretations: Customized Fuzzy Matching, leading to a 10% reduction in user queries.
- Automated In-Car HU Testing: Led development of a ResNet-50 powered image classification system for Mercedes-Benz, reducing in-car HU testing time by 70% and achieving an F1 score in the high 0.8 - 0.9 ranges for UI element recognition. Thus, Streamlining the testing process and facilitated faster bug detection, leading to a more robust and user-friendly in-car app experience
- Led Mercedes-Benz Infotainment Data Parsing (Python): Built Python parser demonstrating strong skills in data structures and algorithms, achieving 80% UDC processing efficiency.
- Generative AI Research: Spearheaded research on Generative AI, exploring methods like Retrieval Augmented Generation (RAG), Hyde, RAPTOR. Utilized LLMs like OpenAIs GPT-3, Claude, AWS Bedrock and Gemini for POC applications with 82% retrieval accuracy.
Data Scientist
DeloitteDec, 2020 - Apr, 2021 4 months
- Reduced Legal Transcript Costs: Contributed to a real-time speech recognition system for legal hearings using Deep Learning, potentially saving the company ~$500,000 annually.
- Improved Speech Recognition Accuracy by 12%: Optimized a deep learning model for legal speech data through hyperparameter tuning, achieving an F1-score of 88% (precision & recall metric). This resulted in a significant 12% improvement in speech recognition accuracy compared to the baseline model.
- Formulated Transcript Clarity with Speaker Diarization: Introduced speaker diarization as a potential solution to differentiate between speakers within legal hearings. This approach is projected to improve transcript clarity by up to 20%, making transcripts more organized and searchable.
Achievements
- • Exemplary Performance: Recognized as a "Star Performer" for consistently exceeding established company benchmarks by 40%. • Mentorship and Talent Acquisition: Successfully trained and mentored over 15 individuals, and Actively participated in hiring for senior positions (T7/T8/T9s) and fresher, contributing to attracting top talent for company growth. • Open-Source Advocate: Made 4 notable contributions to popular Machine Learning libraries like Keras, TensorFlow, and OpenAI Whisper, actively promoting collaborative development within the Machine Learning community. - Speaker for Google , Conscious Algorithms : A Talk on AI Safety.
Major Projects
8Projects
QuantumAR
Developed an Augmented Reality application with OpenCV and Python for real-time feature matching and dynamic object replacement.
Noah
Designed a smart chat/recommendation bot using FastAPI, Python, Dialogflow, MySQL, and NLP for a custom shopping site.
AirFlow
Developed an interactive gesture-based Air Canvas using ML, Mediapipe, and OpenCV for tracking hand movements and recognizing gestures.
Notii-fy
Feb, 2024 - Mar, 2024 1 month
Engineered Notiify, an ad-free, local music player application inspired by Spotify. Utilizing Python and speech recognition, enabling hands-free music control by allowing users to voice-activate song playback. Attained an average voice command response time of 0.88 seconds.
OpinionSense
Jan, 2024 - Feb, 2024 1 month
Coded and implemented OpinionSense, a sentiment analysis system for reviews using a Recurrent Neural Network (RNN) from scratch. Achieved an F1-score of 0.85.
Cropcure_AI
Jan, 2024 - Feb, 2024 1 month
Created CropCure AI, a Flask application leveraging Deep Learning for real-time Blythe disease detection in potato leaves. Gained an F1-score of 88%
paperScribe
Jan, 2024 - Feb, 2024 1 month
Architected PaperScribe, a RAG-based AI using GPT-3 for document understanding and interactive exploration. Leveraging Streamlit, PaperScribe accomplished 92% accuracy in document retrieval tasks based on user queries, facilitating efficient information access from PDFs, articles, and linked content.
MK_LLM
Jan, 2024 - Feb, 2024 1 month
- Devised a bigram character-based architecture, achieving a competitive F1-score of 87.2 on a held-out validation set.
- Trained MK-LLM on a subset of OpenWebText, a massive text dataset comparable to GPT-2 training data.
Education
BE Computer Science
SDMCET, Dharwad
Certifications
Machine Learning
DeepLearning.AI- Stanford (Apr, 2024)