PyCon US gives the community a chance to come together and learn about what’s new and interesting about the Python language and the seemingly infinite variety of problems that can be solved with a few (or a few thousand) lines of Python code. For entrepreneurial Pythonistas, Startup Row at PyCon US presents a unique opportunity for startup companies to connect directly with the developer community they’re building for.
Kicked off in 2011, Startup Row at PyCon US gives early-stage startups access to the best of what PyCon US has to offer, including conference passes and booth space, at no cost to their teams. Since its inception, including this year’s batch, well over 150 companies have been featured on Startup Row, and there’s a good chance you are familiar with some of their products and projects. Pandas, Modin, Codon, Ludwig, Horovod, SLSA, and dozens of other open-source tools were built or commercialized by companies featured on Startup Row at PyCon US.
Think of Startup Row at PyCon US as a peek into the future of the Python software ecosystem. And with that, we’re pleased to introduce the 2025 batch!
The Startup Row 2025 Lineup
AgentOps
Building an AI agent that works is only half the battle; seeing why it fails, how much it costs, and whether it’s about to go rogue is the other half. AgentOps gives developers that missing x-ray vision. Drop a two-line SDK into your code and every run is captured as a “session” complete with step-by-step waterfalls, prompt/response pairs, cost and token metrics, and security flags—instantly viewable in a web dashboard.
The idea was born at a San Francisco hackathon, where co-founders Alex Reibman and Adam Silverman discovered that their agent-debugging tools were more popular than the agents themselves. They turned those internal tools into AgentOps, raised a $2.6 million pre-seed led by 645 Ventures and Afore Capital in August 2024, and now give thousands of AI engineers a live dashboard that replays every agent step, surfaces exact cost and latency metrics, and enforces benchmark-driven safety checks—all from a two-line SDK.
Open-sourced under an MIT license, the project has already racked up 4.4k GitHub stars and integrates out-of-the-box with OpenAI Agents SDK, CrewAI, LangChain, AutoGen and dozens of other frameworks. With observability handled, AgentOps wants to be to autonomous agents what Datadog is to micro-services: the layer that makes ambitious agent stacks safe enough for production—and cheap enough to keep running.
AllHands AI
Agentic coding went from a theoretical possibility to something seemingly overnight, and All Hands AI’s open-source platform OpenHands is one of the reasons why. Written in Python (with a JavaScript front-end), OpenHands lets an AI developer do everything a human can: edit repositories, run shell commands, browse the web, call APIs—even lift snippets straight from Stack Overflow—and then roll it all back into a commit you can review and merge.Since its first README-only commit just 14 months ago, the project has snowballed into 54k-plus GitHub stars and 6k forks, backed by a community of roughly 310 contributors and counting. The momentum helped the team close a $5 million seed round led by Menlo Ventures last September, giving the ten-person startup runway to layer commercial tooling on top of its permissively-licensed core.
“About six months ago it finally clicked—I now write about 95% of my own code with agents,” says co-founder and chief scientist Graham Neubig, an associate professor at Carnegie Mellon who shipped the project’s first lines before Robert Brennan—now CEO—joined the project and built a globally-distributed team to scale it up. Neubig credits the early decision to ship a “non-functional prototype” and build in public for catalyzing the contributor base; today, community members maintain everything from Windows support to protocol bridges while swapping LLM benchmarks daily in the project’s Slack.
OpenHands has evolved from a weekend proof-of-concept into a community-driven framework that now aims for production-grade reliability as an open alternative to proprietary code agents. Weekly releases focus on reproducible debugging, cost control, and enterprise safeguards, and contributors are already using the system to generate and review real pull requests across a growing set of Python projects.
DiffStudio
Product photos tell a story, but DiffStudio wants to let shoppers walk around that story. The North-Jersey startup is building a camera-agnostic “inverse graphics” pipeline that ingests a handful of ordinary 2-D shots or video and returns a fully-textured, web-ready 3-D model that drops into any product page. The goal is simple: turn scrolling into spinning, pinching, and zooming—and watch engagement and conversions rise.Founder Naga Karumuri just recently formed the company in December, after months of hacking on the latest developments in Gaussian splatting and differentiable rendering. “You upload a batch of images, and our model hands you a compressed asset—think megabytes, not gigabytes—that Shopify can serve instantly,” Karumuri explained. A companion mobile app will let merchants scan products on the fly, while a web dashboard handles cloud processing and one-click embeds.
DiffStudio’s beachhead market is small- and mid-sized Shopify sellers, and blue-chip retailers are already circling. “In casual chats we’ve had interest from brands like Adidas and Michael Kors,” Karumuri noted, hinting at an eventual move up-market once the self-service MVP launches. Compression and quality are the differentiators: where existing tools like Polycam focus on hobbyist scans or LiDAR-assisted captures, DiffStudio is chasing photo-real fidelity with file sizes that won’t tank page speed. The project’s GitHub repositories showcase early demos and the startup’s open-source commitment.
The team is still lean—Karumuri plus a collaborator—but the vision is outsized: make 3-D product “digital twins” as easy to generate as a product photo set. Or, as their LinkedIn banner puts it, “Splat your products into 3D glory.”
Fabi.ai
Business users shouldn’t have to ping the data team for every ad-hoc question—and data scientists shouldn’t spend half their day writing the same queries on repeat. Fabi.ai positions itself as the AI “side-kick” that lets both camps meet in the middle: a web notebook where natural-language prompts, SQL, Python, and no-code building blocks live side-by-side, with generative agents filling in (and explaining) 90 % of the boilerplate.Founded in 2023 and headquartered in the San-Francisco Bay Area, the six-person team works face-to-face in San Mateo to iterate quickly on the product. CEO Marc Dupuis ran embedded analytics at revenue-ops unicorn Clari before teaming up again with former colleague Lei Tang (now CTO) to “let vibe-coders do 95 % of their own analysis” while still giving experts an easy way to supervise the last mile.
Eniac Ventures and Outlander VC co-led a $3 million seed round in July 2023 to bring Fabi.ai’s collaborative notebook to market. Early customers already range from fast growing startups to established e-commerce brands.
With BI dashboards stuck on the what and legacy notebooks siloed on individual laptops, Fabi.ai is betting that a cloud-native, agent-augmented workspace is the missing link—and it’s inviting the Python community to kick the tires (and write fewer queries) at PyCon US.
Gooey.ai
Most no-code AI builders stop at slick demos; Gooey.ai is obsessed with what happens after the hype, when a multilingual chatbot has to work for a Kenyan farmer with a 2 G signal or a frontline nurse switching between English and Kannada. The open-source, low-code platform stitches together the “best of private and public AI” into reusable workflows—text, speech, vision and RAG pipelines you can fork, remix and ship to WhatsApp, SMS, Slack or the web from a single dashboard. One billing account, one-click deploy.Founders Sean Blagsvedt (ex-Microsoft Research, founder of Indian job-matching startup Babajob), Archana Prasad (artist-turned-social-tech entrepreneur), and CTO Dev Aggarwal split their time between Seattle and Bangalore and run the company under the umbrella of Dara Network. Their thesis: impactful AI needs to be both affordable and local—so Gooey bakes in speech recognition for low-resource languages, translation APIs like India’s Bhashini, and zero-data-retention options for NGOs handling sensitive chats.
Real-world traction is already visible. An agronomy WhatsApp bot built on Gooey reached “tens of thousands of farmers in Kenya, India, Ethiopia and Rwanda,” delivering accurate, objective answers with page-level citations. The platform’s copilot builder now supports the latest GPT-4o, Llama 3, Claude, Gemini and Mistral models; integrates OCR, vision and text-to-speech; and ships bulk evaluation harnesses so teams can test new prompts before they hit production.
To seed more grassroots projects, Gooey recently launched a Workflow Accelerator with funding from The Rockefeller Foundation, covering model and SMS costs for six NGOs and open-sourcing every workflow that emerges. If you’re looking to take an AI pilot from “cool demo” to “24/7 field tool,” Gooey.ai wants to be the glue—and the infra—you won’t outgrow.
GripTape AI
Enterprise AI teams love the idea of autonomous agents, but hate the roulette wheel of prompt-only code. Griptape steps in with a Python framework that keeps creativity where it belongs—inside LLM calls—while wrapping every outside step in predictable, testable software patterns. Agents, sequential pipelines, and parallel workflows are first-class “Structures”; memory, rulesets, and observability are plug-in Drivers; and an “Off-Prompt” mechanism pushes sensitive or bulky data out of the prompt for lower cost and higher security.The project launched in early 2023 and has already gathered ≈2.3 k GitHub stars and an active Discord community. Adoption accelerated after co-founders Kyle Roche and Vasily Vasinov—both former AWS leaders—closed a $12.5 million Seed Round in September 2023 led by Seattle’s FUSE and Acequia Capital. The fresh capital funds Griptape Cloud, a fully managed runtime that hosts ETL pipelines, hybrid vector knowledge bases, and structure executions while piping metrics to whatever monitoring stack a Fortune 500 already uses.
Under the Apache-2.0 license, developers can start locally, swap between OpenAI, Bedrock or Anthropic drivers, and graduate to the cloud only when they need auto-scaling or policy enforcement. In short, Griptape aims to be the Django of agentic AI: batteries-included, prod-ready, and with enough guardrails that even the compliance team can sleep at night.
Griptape also recently launched Griptape Nodes, an intuitive, drag-and-drop interface where designers, artists and other creative professionals can create advanced creative pipelines using graphs, nodes, and flowcharts to exploit state-of-the-art image generation and image processing models, together with more “traditional” large language models.
MLJAR
Most AutoML platforms lock you into a browser tab and someone else’s GPU cluster. MLJAR takes the opposite approach: everything runs locally, yet you still get the “train, explain, and deploy” cycle in a single click.The Polish-based project began in 2016, when founder Piotr Płoński—fresh from a PhD spent building models for physicists, bioinformaticians, and telecom giants—decided he was tired of rewriting the same pipelines over and over. Impatience, not laziness, pushed him to automate the entire workflow.
Today the three-person team (Piotr, his co-founder wife, and a close friend) maintains a fully open-source stack. The flagship MLJAR-AutoML package handles feature engineering, hyper-parameter search, and rich Markdown reports; Mercury turns any Jupyter notebook into a shareable web app or dashboard with a sprinkle of widgets; and the brand-new MLJAR Studio Desktop app bundles its own Python environment, point-and-click “code recipes,” an integrated GPT-4 assistant, and a one-button Share that converts a notebook into a live web application.
Open source is more than a distribution strategy—it’s a trust signal. One recognisable enterprise adopted the package under an MIT license and then contracted the team for advanced features such as fairness-aware training. Revenue is a side effect; the primary goal is software that makes data science faster, friendlier, and fully under the user’s control.
If you’ve ever wished Streamlit met AutoML—and ran natively on your laptop—swing by the MLJAR booth on Startup Row at PyCon US and take Studio for a spin.
Ragas
Seemingly everyone is building RAG pipelines, but almost no one is measuring them. Ragas sets out to be “pytest for Retrieval-Augmented Generation,” bundling ready-made metrics—context recall, faithfulness, answer relevancy—and auto-generated test sets so teams can turn vibe checks into repeatable CI tests. Drop the library into LangChain, LlamaIndex, or plain-Python code and Ragas spits out a single “Ragas Score” (plus per-metric drill-downs) that tracks whether your latest prompt tweak fixed accuracy or broke it.The project landed a shout-out during OpenAI’s Dev Day and has since snowballed to 9.1 k GitHub stars and 900+ forks, with more than 80 external contributors. In production it now processes ~5 million evaluations a month for engineers at AWS, Microsoft, Databricks, and Moody’s—a number growing 70 % month-over-month.
Co-founders Jithin James (early engineer at BentoML) and Shahul ES (Kaggle Grandmaster, core contributor to Open-Assistant) met at college, hacked on open-source together for years, and entered Y Combinator’s W24 batch to turn their weekend project into a commercial platform. Their plan: keep the core evaluator MIT-licensed while DG Labs, the commercial arm, layers team dashboards, experiment tracking, and dataset management on top—so every product squad can ship RAG updates with CI-style confidence.
Thank You’s and Acknowledgements
There are far too many stakeholders in the ongoing success of Startup Row at PyCon US to name individually, but this program would not be possible without the following folks:- The Python Software Foundation, for its continued support of this little corner of PyCon US.
- The PSF Sponsorship team, for managing the logistics of getting everyone registered and set up for success
- Startup Row co-organizers, Jason D. Rowley (p.s. hey, that's me!) and and collaborator, Shea Tate-Di Donna, whose first experience with the Python community was presenting her company, Zana, on Startup Row at PyCon US 2015.
- Startup Row alumni companies that come back as paid sponsors at PyCon US. Shoutouts to Anvil (SR’17), Chainguard (SR’22), and Dagster (SR’21), whose support helps make Startup Row at PyCon US possible.
- To all startup founders who filled out the (mercifully brief) application. To those that did not get a spot this year, we appreciate your time and attention. To those that did: a hearty congratulations.
- To the selection committee, for accomplishing the difficult task of evaluating and scoring applications.
2026 will mark 15 years of Startup Row at PyCon US. We can't wait to see what's next in the Python software ecosystem. 🐍
Comments