Our vfs is also pretty powerful too, though: it is all backed by postgres then projected into files/directories for our agents. They get basic grep etc but also optimized fts tools for bm25, jq, and preview tools that show representative slices of large documents. All on top of Pydantic AI.
We use both a virtual file system and RAG — they each excel in different areas. The trick with RAG is the quality of data: we use an LLM to chunk into semantically cohesive sections, as well as generate metadata (including fact triples and links to other related chunks in the document) for every chunk as well as the document as a whole. We use voyage contextual embeddings to then embed each chunk with the document and chunk metadata. Works incredibly well. At retrieval time the agent can follow chunk links if needed, as well as analyze the raw file in the vfs. High quality instruction based reranking helps a lot too! We are often looking over 10s of thousands of documents and it’d be very inefficient to have our agents analyze just the vfs without rag.
Hiring for a full-time, 100% remote role (entire company is remote) for a Staff Software Engineer, working on the ShiftUp platform stack (Python, Pydantic AI, GCP, Temporal Workflows, Turbopuffer, Postgres). You will be one of the first five engineering hires, and you will have the ability to grow and shape your own role.
We need someone with solid experience designing and implementing systems with many moving parts, who can easily shift between systems design, writing and reviewing code (or, let's face it, supervising Codex/Cursor/Claude Code), working with the team to design features, and talking to customers to gain feedback and insights.
You'll be working to build all kinds of agents, workflows, and other components to support our features. Our research agents are truly deep, sometimes spending upwards of an hour searching and analyzing information for a single prospect or strategy. Our RAG pipeline is focused on quality, leveraging LLM chunking, metadata and graph generation, and contextual embeddings.
If this sounds like fun, let's chat: you won't be asked to regurgitate leetcode solutions, we'll simply have a conversation and see if there's a mutual fit. Our interview process is lightweight, respectful and moves fast. Competitive salary, equity and full benefits.
About the company: ShiftUp is a new kind of intelligence-driven sales platform that identifies emerging demand and drives revenue. Built from scratch to leverage LLM-driven agentic workflows, ShiftUp deeply researches prospects, identifies the highest-value opportunities, and provides targeted sales strategies to execute on them.
ShiftUp is delivered as a native app on Salesforce AppExchange, providing our platform to sellers where they already are. Our Salesforce app talks to a cloud-native stack that uses Python services, durable workflows, and SQL + vector databases.
We're well-funded, having recently closed a $3M seed round. We're building out a small team of highly talented, well-rounded engineers to deliver on our vision. Our company is 100% remote, and we use SoWork to help build our culture. As a small startup, everyone is highly motivated and focused on moving quickly, but we also place importance on a sustainable work-life balance so that you can continue to do your best work.
I really miss the everything is editable panel, it felt like a superpower. There’s a bit of a learning curve, but after it’s amazing and everything else feels limited.
If anyone was trying it out, we just had an issue with one of our backend services hitting a usage limit. All fixed now so if you had a failed refresh please try again.
OP here - feel free to ask any questions. We are still trying to figure out what direction to take Jumprun in from here and would love your feedback and ideas. Thanks!
We built Jumprun. You can use it to research and analyze data sources, and it'll produce beautiful canvases with tables, charts, videos, maps, etc. We're working on automations so you can setup natural language trigger conditions that execute actions.
We built it in Kotlin with Ktor server, htmx and tailwind. It uses a mixture of models, including gpt4-turbo, gpt4-vision and gemini-pro-vision. It's deployed using Kamal on bare metal.
Very easy! You just need to ground it with the right data sources. Eg. you can add a web search datasource for "oculus reviews", and then add a table component to your canvas, instructing it to provide a comparative analysis.
The source material doesn't need to provide a comparative analysis itself: as long as it finds data detailing Vision Pro specs (which it already found) and Oculus specs it'll be able to do the analysis.
Canvases also have memory and auto-refresh, so once it finds the specs it can remember them across refreshes/updates, but also incorporate any new information it finds.
Just a little more context: Jumprun let's you connect different data sources (like web searches/pages, APIs, X, Youtube videos, Notion etc) and use LLMs to analyze and visualize the data.
We support rich components like tables, timeseries, charts and maps. We're working on automations at the moment so that you can provide natural language conditions that trigger actions (like sending you an email or changing updating a page in Notion).
Our long-term vision is that canvases become more interactive and interconnected, so that you end up building mini applications without it feeling like you're using a low-code app builder.