SUMMARY

AI/ML Engineer with proficiency in Large Language Models, both with fine-tuning & prompting and CI/CD processes. Possessing a strong software engineering background that includes writing typesafe code with pydantic/mypy, unit & integration tests, and creating GraphQL APIs. Recently I have been working on:

Databricks to improve a data analytics pipeline that generates open & reply rate statistics for our product dashboard. I like how they make it very simple to connect different data providers with tasks, which can also be used as input for various types of ML experiments. Also easily log these to MLflow for observability & reproducibility.
Llama.cpp & LlamaIndex python package and its capabilities. It gives a much better flexibility compared to the APIs. I could easily control the output format with a few lines of code. See my latest medium blog here.

PRIMARY WORK EXPERIENCE

LEADIQ (San Francisco, CA)

Tech Stack : PyTorch, GPT, OpenAI API, Strawberry, FastAPI, Flask, Streamlit, Transformers, HuggingFace, Langchain, Gitlab, Linux, AWS, MongoDB, Docker, Datadog, Insomnia, Sisense, Redis, Redshift, Databricks, PySpark, Slack, Jira

Machine Learning Engineer (January 2022 – Present)

Developed personalized prospecting email generator using Large Language Models (LLMs) that helped users to reach out to >5x more prospects with competitive open&reply rates
Improved the quality of emails generated by utilizing automated user feedback to either fine-tune or improve prompting of our LLMs
Utilized with many different fine-tuning techniques to increase performance including upside-down RL / decision transformer style, weighting of user feedback data, various sampling strategies to reduce entropy collapse and improve variety
Improved stability of Databricks pipeline that computes open & reply rate statistics daily.
Created an account research application using Streamlit and LLMs, which users can search for accounts and leads and get a set of insights powered by many APIs that are then processed with LLMs to generate a summary
Created a chatbot application, using Langchain and Streamlit, helping users identify pain points of a prospect and then write an email using the user instructions, utilizing many APIs as tools
Developed usage and performance dashboards using Sisense, a product to build intelligent analytics, so our customers can monitor the statistics of the emails that we generate for them.
Trained a job title seniority & function classifier using bi-directional LSTMs, which replaced the old rule-based model and is used widely in production which achieved >90% classification accuracy among 6 seniority classes and >70% accuracy among 50 job functions that was used by the most central services in the company
Trained a job title similarity scorer with a Siamese network achieving >95% classification accuracy
Designed and developed APIs with GraphQL, wrote typesafe code with pydantic/mypy, wrote integration tests and unit tests, reviewed merge requests, inspected and resolved errors from datadog logs and customer complaints, collaborated with project manager and other team members
Refactored existing code for performance; for example re-wrote a code snippet that finds the ticker of a company given its domain, helping us increase our precision by absolute 15%

SESTEK (Istanbul, Turkey)

Tech Stack : PyTorch, BERT, LSTM, FastAPI, Transformers, HuggingFace, Linux, Docker, Azure DevOps, Azure Cloud

Senior AI Research Engineer (March 2020 – December 2021)

Conducted comprehensive literature reviews and wrote reports to ensure products stayed up-to-date
Organized regular meetings with auditors for R&D projects to ensure the continuation of the investments
Created a parent Python project for Machine Learning (ML) model pipeline to automate the training and deployment process which evolved into SestekAI, a product helping customers train and serve their models in their local environments

Research and Development Engineer (March 2017 – March 2020)

Trained intent detection and FAQ text classifier models using PyTorch, Hugging Face libraries, LSTM, and Google BERT architectures, delivering 85% intent detection accuracy to the chatbot application of one of Turkey’s largest banks
Developed a sequence-to-sequence LSTM model that efficiently transformed textual representations of numeric values such as addresses, dates, and times into formatted forms, leading to a publication on this subject
Took part in every stage of the ML Pipeline: data cleansing, training, evaluation, testing on real customer data, and deploying trained models as REST services using Python

EDUCATION

BOGAZICI UNIVERSITY (Istanbul, Turkey)

2020 - Master of Science: Electrical and Electronics Engineering (GPA: 3.69)

Research Focus: Automated Response Generation for Corporate Chatbot Systems
Specialized Courses: Pattern Recognition, Speech Processing, ML, Statistical Signal Analysis, Social Semantic Web

2017 - Bachelor of Science: Electrical and Electronics Engineering (GPA: 3.28)

Core Courses: Probability, Matrix Theory, Signal Processing, Introduction to Image Processing

TECHNOLOGIES

Python : PyTorch, Langchain, OpenAI API, Strawberry, FastAPI, Flask, Streamlit, Chainlit, Numpy, Pandas, Pydantic
AI/ML : GPT, Transformers, LSTM, BERT, Hugging Face, Llama, LlamaIndex, Mlflow
Data Analytics : Databricks, PySpark, Redshift
CI/CD : Gitlab, Github, Docker, Datadog, Insomnia, Redis, Slack, Jira, AWS, MongoDB, Azure DevOps
Other : Linux, Sisense, GraphQL

PUBLICATIONS

A. S. Güser, M. Erden and M. L. Arslan, “Semi-Automatic Formatting of Spelled Out Numbers” 2019 27th Signal Processing and Communications Applications Conference (SIU), Sivas, Turkey, 2019

CERTIFICATES

ADDITIONAL

Achievements : Ranked 6th among 1.7m students in the National University Entrance Exam (2012)
Activities : Walking, Literature, and Traveling

Abdullah Şamil Güser