-
Devang
Pagare
linkedin.com/in/devang-pagare
#-
ï
§ github.com/DevangPagare002
Education
B.Tech. Artificial Intelligence (Honors in Data Science) [CGPA - 9.54]
G.H. Raisoni College of Engineering and Management
Jan 2020 – June 2024
Pune, India
Experience
Unicore Aiminds Pvt. Ltd
May 2022 – Nov 2022
Software Intern
Pune, India
• Developed a web-based application for handling and monitoring multiple client-side IoT hardware and sensors.
• Designed the database schema to store all the data from sensors.
• Implemented MQTT protocol and utilized HiveMQ as a reliable broker for seamless connectivity between software and
hardware devices; reduced data transmission latency by 50% and enhanced real-time monitoring capabilities.
• Used the simulator to verify software integrity and connectivity. Created/Updated and implemented protocol and
payload guide.
Winsoft Technologies India Pvt. Ltd.
Aug 2023 – Feb 2024
Project Trainee
Pune, India
• Worked in the R&D department under the Senior Vice President and developed AI solutions for the Finance sector.
• Used Machine learning algorithms like Regression, Random Forest, and Time Series models(ARIMA, ARMA, etc) for
problems like churn prediction, demand prediction, forecasting, fraud detection, etc.
• Finetuned Pythia LLM on Lamini docs to improve output quality. Worked on vector databases like Qdrant, Faiss,
Pinecode, and Chroma. Created the first quantized version of Llama2-7B-Finance model in GGUF format with 16bit,
32bit and 8bit precision quantizations.
• Developed a Q&A bot based on RAG that uses vector database and LLM to answer questions based on existing or
provided knowledge base.
• Developed the AI-powered personal Finance Advisor that daily advises investors about their investments and suggests
recommended actions. It uses LLMs, News APIs, Streamlit, Qdrant vector database, etc.
Emergiq
May 2024 – Present
AI Backend Engineering Intern
Remote
• Developing entire flask backend with multiple endpoints for
app
in
WSGI
production
server
ensuring
robust
Emergiq
§
and scalable architecture. Initially deployed the backend using Google Kubernetes Engine (GKE), later transitioned to
Digital Ocean’s App Platform for improved deployment efficiency.
• Actively leading the backend development and MLOps to drive the project’s success and deliver high-quality results.
Engineered several prompts targeted to different modes of the app, utilizing advanced techniques to improve user
interaction and experience. Developed CI/CD pipeline with GitHub Actions.
• Designed and implemented the app’s database schema using Firestore, optimizing for performance and scalability.
• Utilizing LLMs like Llama3, Mixtral, Gemma, Gemini, and APIs like Groq and Deepinfra. Creating and Managing cron
jobs using Digital Ocean’s functions with triggers to automate various backend processes.
Projects
Finetuned EluthreAI’s Pythia 1.4B/70M LLM | Python, LLM, Lamini dataset
• Finetuned ElutherAI’s Pythia 1.4 billion and 70 Billion models using Transformers, accelerate library on lamini docs.
• 4+ Finetuned models are saved on § Huggingface, they can be reused.
• Models are trained on multiple epochs and each one has a memory footprint of 5.7 GB.
DocGenius(an RAG app) | Python, Langchain, Qdrant, Streamlit, LLM
• This is a prototype of an app that can answer questions on given PDFs which is available on
§ Huggingface spaces.
• This app uses all-MiniLM-L6-v2 embedding algorithm and llama-2-7b-chat.Q3KS.gguf model loaded with the help of
Llama-cpp. It uses the Qdrant vector database to store all the data vectors in the local machine and cloud.
Robo-Advisor | Python, Langchain, Streamlit, yfinance
• This is a personalized AI-powered app that can advise people about their investments
• It takes the invested company as input and then searches for the latest financial news regarding that company.
• Based on the news article, it suggests what action (buy/sell/hold) to take on the stock and justifies that action.
Technical Skills
Languages: Python, C++, HTML/CSS, JavaScript, SQL
Developer Tools: VS Code, Eclipse, Git, Anaconda Navigator, Pycharm, Visual Studio
Technologies/Frameworks/Libraries: NLP, Huggingface, Langchain, Llamacpp, Ctransformers, llama-cpp-python,
AutoLLM, Auto-GPTQ, Transformers, Prompt Engineering, Accelerate, Chroma, DataStax AstraDB, Qdrant,
BitsandBytes, Linux, Pytorch, Transformers, GitHub, Deep Lake, Lucidchart, Docker, Kubernetes, GKE, DigitalOcean,
Firebase, Firestore, Groq, Deepinfra.
LLMs: GPT2, Pythia-70M, Pythia-1.4B, BERT, RoBERTA-base, T5-flan-large, Zephyr, Mistral, Llama2, Mixtral, Llama3,
Gemini, etc.
