4 ms·
Location: San Francisco, CA Remote: Yes Willing to Relocate: No Programming Languages: Ruby, Scala, Python, Javascript, TypeScript, Haskell Programming Stre
by poeticsilence 3y ago
Location: San Francisco, CA
Remote: Yes
Willing to Relocate: No
Programming Languages: Ruby, Scala, Python, Javascript, TypeScript, Haskell
Programming Strengths: Programming Language Theory, Type Theory, Data Structures and Algorithms, Data Engineering, No-Code, AI&Prompt Engineering
Other Skills: Leading Teams & Managing Projects, Architecting & Implementing Greenfield Solutions, Education & Team Growth, Leading Communication and Coordination Between Stakeholders Across Teams and Organizational Silos
CV: Available Upon Request
Email: poetic.artifice@gmail.com
I am a seasoned (15 years experience) full-stack principal software engineer specializing in the backend. In the past, I've built a number of no- and low-code solutions for stream processing, data model and query generation, form building, workflow systems, and front-end construction for a variety of industries, including those with strong regulatory requirements and complex business rules such as healthcare and fintech.
More recently, I've built a workflow engine for coordinating LLMs and other generative AI processes, and on top of this a platform for minimizing hallucinations and working with documents that are too large for existing context windows.
If you are building a greenfield AI project or adding AI to your existing business and your business is particularly sensitive to the consequences of hallucinations or you need to work with a document corpus that is larger than context windows allow for precise recall within, please don't hesitate to reach out.
- samier-trellis 3y agoHey, I want to encourage you to apply to this position at our legal tech startup: https://jobs.lever.co/trellis/b76fa20e-8bf4-4d6e-a716-200f369f367e https://jobs.lever.co/trellis/b76fa20e-8bf4-4d6e-a716-200f36... We are looking for an AI/ML engineer with experience building products using LLMs. Two of our chief pain-points currently are dealing with documents that are really big and exceed existing context windows and building a "RAG" search that is more than just vector search.