Suggestions
John Damask
Vice President, Data & Systems Engineering at Flagship Pioneering
John Damask is a seasoned professional currently serving as the Vice President of Data & Systems Engineering at Flagship Pioneering, a position he has held since May 2019.1 With over 20 years of experience in delivering value through software and informatics solutions, Damask brings a wealth of expertise to his role.1
Professional Background
Prior to joining Flagship Pioneering, Damask had an impressive career trajectory:
Novartis Institutes for BioMedical Research (NIBR):: He spent eleven years at NIBR, taking on various roles of increasing responsibility. His most recent position there was Global Head of Cloud Computing, where he led the team responsible for all aspects of the corporate cloud platform.1
Consulting Practice:: Earlier in his career, Damask ran his own consulting practice for a decade, creating innovative software and data products across various industries.1
Expertise and Achievements
John Damask's areas of expertise include:
- Leadership and management of global teams
- Value creation and delivery
- IT Service Management and Operations
- Amazon Web Services (AWS)
- Data and software architecture
- Coding in multiple languages
- Bioinformatics with a focus on proteomics
- IoT and Bluetooth Low Energy technologies
Some of his notable achievements include:
- Building global automated platforms for high-throughput labs at NIBR, resulting in savings of millions of dollars and decades of manual effort
- Developing integrated data warehouses that allowed for rapid, in silico explorations of diverse topics including chemical biology, proteomics, transcriptomics, and proteogenomics1
- Receiving an NIH grant to develop a decision support system for proteomic researchers1
Current Role at Flagship Pioneering
At Flagship Pioneering, Damask's responsibilities involve:
- Providing strategic leadership around the use of data and technology within Flagship and across its portfolio
- Overseeing cloud and scientific computing, architecture, engineering, data management, automated operations, and governance1
Education
John Damask holds a B.A. in Musicology from New York University.1
Recent Activities
Damask remains active in his field, participating in industry events and initiatives:
- He attended the Amazon Web Services (AWS) Life Science Symposium in Boston in May 2024.4
- In June 2024, he was involved in Flagship Pioneering's first AI Engineering Hackathon, hosted in collaboration with AWS.5
- As of September 2024, Damask's team at Flagship Pioneering was expanding, with openings for Director/Senior Director of Digital Science Solutions.3
John Damask's LinkedIn username is johndamask, where he describes himself as someone who loves "to create lasting value from scratch and build exciting, healthy" teams.2
Highlights
I'm a fan of Shreya Shankar's work. She's a big thinker, asute problem prioritizer, and clear communicator.
In this blog, https://lnkd.in/eDJWZbPA Shreya presents an idealized model for building "self-improving LLM applications" in which responses are continuously checked to ensure they stay on point over time.
Evaluating what LLMs produce is an extremely important part of putting an AI app in production because once it's out there, everyone assumes it works as intended, i.e. perfect and complete responses every time. But if you asked an LLM to swear to tell the truth, the whole truth and nothing but the truth, they'd dodge the question about go about their business. This can lead to a loss of trust.
Being aware of these problems is important for building apps as well as becoming an educated customer. Thanks Shreya
Flagship Pioneering threw our first AI Engineering Hackathon this week, hosted by our good friends at Amazon Web Services (AWS). Our theme, "Chat With Your Database," demonstrated how LLMs enable text-to-SQL translation, making data sources easily accessible to non-data scientists.
Teams from several of our portfolio companies joined and whipped up some fantastic PoC applications. The stack that ruled the day was Streamlit, LangChain and AWS Bedrock and the preferred models were Claude and GPT4o.
Some learnings:
- It's essential to provide the LLM with the table definitions of your database - some teams found that passing these via the System Message yielded better results than in the User Message.
- The LLM tends to make better decisions when textual descriptions for tables and columns are included.
- Model Temperature is thought to play a role but this wasn't thoroughly tested. It's presumed that a Temperature of 0 will produce more consistent results than higher temps
- Getting the prompts right is tricky and requires lots of testing. It's not clear whether techniques like chain of thought or few-shot perform better than zero-shot
- One team found that instructing the LLM to pay particular attention to numbers in queries gave better results. That is, without this guidance something like "give me the top 4 targets" may produce SQL that ignores the top 4 filter
It was a great day that showcased the power of the Flagship community
