| style | title | social | ||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
serenity |
Nidhish Shah |
|
- Engineered automated, no-code training and evaluation pipelines for large language models (LLMs), transitiong the team to fully open-source frameworks from proprietary cloud dependencies.
- This initiative has faciliated day-one training and evaluation of newly released LLMs, significantly enhacing operational efficiency.
- Led the fine-tuning and deployment of a model for data extraction, achieving over 98% reduction in operational expenses from $60,000 to under $1,000 monthly, while delivering superior performance compared to existing proprietary closed-source models.
- Technologies: Large Language Models, Python (PyTorch, HF Transformers), Docker, AWS.
- Developed a fraud-detection system that detected anomalies of up to 10M USD in Azure consumption for clients across Western Europe.
- Integrated image segmentation model into a knowledge mining system to extract, name, and index chemical structure images, eliminating search time for un-indexable research documents for a top-15 bioscience company in the Netherlands.
- Designed and trained a robotic arm simulation using the MuJoCo Engine and Microsoft Project Bonsai to handle soft bodies with precision.
- Technologies: Python, Azure, SQL, Computer Vision, Reinforcement Learning
- Integrated internal build system to automate repository creation and user authorization, resulting in a one-click solution that will save approximately 800 SWE hours annually and enable mass adoption of our framework by 200+ internal teams.
- Created a script to automate the generation of config-files for stress-testing, reducing deployment times of tests by approximately 75%.
- Developed a customer chat-bot using Amazon Kendra & Lex to query internal wikis, allowing instantaneous customer response time for FAQs about our framework.
- Technologies: Java (backend), Python, AWS, Natural Language Processing
- Modelled a reinforcement learning (RL) environment in Unity to mimic a game of laser-tag.
- Utilised the ML-Agents framework and the stabe-baselines3 PPO implementation to train the agents.
- Implemented fundamental reinforcement learning algorithms from scratch like DQN, PPO, and SAC using PyTorch.
- Matched the performance of standard libraries like CleanRL & StableBaselines3 on the MuJoCo Benchmark.
- Winner (3/2000+) of Andrew Ng's Data-Centric AI competition. Presented solution at the NeurIPS Datasets & Benchmark track. 2021
- Academic Excellence Award by the Dutch Society of Sciences (KHMW) for rank 1/350 at TU/e. 2021
- Current GPA: 8.9/10 (US 4.0/4.0).
- Thesis: Exploring the Landscape of Differentiable Architecture Search with Reinforcement Learning under Dr. Joaquin Vanschoren.
- Teaching Assistant for the courses: Introduction To Programming and Data Structures & Algorithms.
- Programming: Python (PyTorch, HF Transformers, Scikit-Learn, Matplotlib), Java, SQL, HTML/CSS, Spark
- Languages: English (native), Dutch (A2)