commit beb9ce4ba7244fb718586320c4b5a5f2fea40c77
parent b4b1f94cf59e7e27fcb72e90ecd1de31e99e1cc5
Author: Silas Brack <silasbrack@gmail.com>
Date: Tue, 13 Aug 2024 19:12:35 +0200
Small changes
Diffstat:
1 file changed, 7 insertions(+), 6 deletions(-)
diff --git a/main.typ b/main.typ
@@ -76,9 +76,9 @@
],
)
#text(size: 9pt)[
- - Designed a real-time recommendation system architecture by adapting industry best practices to our internal data and business objectives, taking into consideration integration of A/B testing, performance/latency, and flexibility to future changes. Presented this architecture to stakeholders and other developers. Implemented this recommender system, training representation learning models, neural networks, and other classifiers and rankers. Designed data processing pipeline (combining translation, text embeddings, and tagging) for preparing content before feeding into recommender engine and serving to our 1M clients. Trained neural networks
- - Developed proof-of-concept of hybrid chatbot and search functionality for searching for / interacting with news and other content, stocks, and much more. Presented and demoed to business stakeholders to provide guidance on how to extend this system for use in in-platform search engine and digital sales trader.
- - Re-designed chatbot combining simple open-source embeddings models and other methods, yielding a 170\% improvement in question-answering accuracy, decreasing the number of required yearly agent interactions by around #format-thousands(20), and saving around \$#format-thousands(150) in yearly agent costs. // % and indirectly causing an increase in KPIs for the client service team.
+ - Designed a real-time recommendation system architecture by adapting industry best practices to our internal data and business objectives, taking into consideration integration of A/B testing, performance/latency, and flexibility to future changes. Presented this architecture to stakeholders and other developers. Designed data preprocessing pipeline (translation, text embeddings, tagging, duplicate detection, etc) for preparing recommended content. Implemented and deployed the recommender system by training neural networks for representation learning and reranking. Tested the recommender engine for quality and latency, achieving a latency of under 100 ms.
+ - Developed proof-of-concept of hybrid chatbot and search functionality for searching for / interacting with news and other content, stocks, and much more. Presented and demoed to business stakeholders to extend this system for use in in-platform search engine and digital sales trader.
+ - Re-designed chatbot combining simple open-source embeddings models and other NLP methods, yielding a 170\% improvement in question-answering accuracy, decreasing the number of required yearly agent interactions by around #format-thousands(20), and saving around \$#format-thousands(150) in yearly agent costs.
- Performed core MLOps tasks, such as standardising data science workflows and development environments with cookiecutter templates, extending Docker environments for development, unifying linting and formatting approaches, etc. Led code reviews and mentored/onboarded new members of the team with understanding our coding practices and infrastructure, focusing primarily on testability.
]
@@ -97,7 +97,7 @@
],
)
#text(size: 9pt)[
- - Developed a machine learning model for finding similar stocks, resulting in our team winning an internal Hackathon competition for improving client experience. This model was then packaged, containerised in Docker, orchestrated as a microservice in Kubernetes and exposed via FastAPI/OpenAPI, yielding a response time of under 10 ms. As of March 2023, this system has been integrated into the official Saxo Bank trader and investor platforms, which is used by nearly one million active clients. As of April 2024, the tool sits at around #format-thousands(70) monthly interactions.
+ - Developed a machine learning model for finding stocks "Others were interested in", resulting in our team winning an internal Hackathon competition for improving client experience. Packaged, containerised (Docker), and deployed this model in Kubernetes, exposing it via FastAPI and OpenAPI, achieving a response time of under 10 ms. As of March 2023, this system has been integrated into the official Saxo Bank trader and investor platforms, which is used by nearly one million active clients. As of April 2024, the tool sits at around #format-thousands(70) monthly interactions.
- Upgraded existing LightGBM forecasting models (churn, revenue, logins, trades, etc.) to train in parallel using Python, Spark (via Databricks) and Synapse ML, increasing training dataset from tens to hundreds of gigabytes of data, sitting in Kafka, Azure Data Lake, and Microsoft SQL Server databases. Scheduled models to train automatically using Airflow.
]
@@ -120,7 +120,8 @@
],
)
#text(size: 9pt)[
- - asdfasdf
+ - Lorem ipsum odor amet, consectetuer adipiscing elit. Urna dignissim fusce leo risus commodo platea ipsum magnis. Id porttitor posuere quis auctor magnis cras suspendisse aliquam. Curabitur dui ante ultricies ultricies nisi ultricies taciti; finibus curabitur.
+ - Duis eu suspendisse morbi eget molestie pulvinar adipiscing a urna. Dignissim faucibus diam blandit himenaeos hendrerit cras eu. Sem ex sapien quis; sodales orci curabitur ante ultricies at. Suscipit nisi sociosqu ipsum mattis urna consequat.
]
#grid(
@@ -196,7 +197,7 @@
*Models*
],
align(left)[
- Python, R, SQL, C\#, Scala, Gleam \
+ Python, R, SQL, C\#, Scala, Gleam / Elixir / Erlang \
Spark, Databricks, Docker, Kubernetes, Airflow, git, Azure Cloud, Terraform, Kafka \
Gradient-boosted trees (LightGBM), neural networks (PyTorch, JAX), Word2Vec, text embeddings, LLMs
],
