We built Last Lab to turn papers, videos, playlists, and other technical resources into verified, running Jupyter labs without the usual setup and guesswork.
We loved how Coursera combines video, in-context prompts, quizzes, and graded exercises. We wondered: what if that same active experience could be created for any reproducible material, from a 3Blue1Brown video or an Andrej Karpathy playlist to a paper, PDF, or repository?
As we built the system, we saw a much larger opportunity. Reproducing a research paper can take weeks or months. A system that extracts the source correctly, resolves its artifacts, and delivers a working implementation can give researchers and technical teams that time back.
See how researchers use Last LabLast Lab turns technical knowledge into a verified Jupyter lab, then surrounds it with what you need to go deeper: source-grounded explanations, extracted artifacts in Lens, active exercises, an AI assistant, and ready-to-use compute along with tools for active exploration.
General-purpose coding agents such as Claude and Cursor can help write code, but reproducibility starts before coding. The source must be extracted correctly, including equations, tables, diagrams, citations, and artifacts, and understood as one connected body of context.
This is a harness problem, not an intelligence problem. Last Lab is designed around that workflow: it extracts the resource, resolves dependencies and artifacts, configures the environment, and executes the result before delivery.
The lab then runs in the browser with CPU and a generous free GPU tier. You do not need to move the output elsewhere, rebuild the environment, or watch an agent reinvent the setup each time. Small points of friction disappear, so you can stay with the work.
The full product details and workflows live where they belong.
A small team focused on making technical knowledge executable.
Taha's been building B2B enterprise AI platforms from 0-1 for over 3 years, covering agent systems, evaluation, and the infrastructure underneath them. At Spheresmith, he owned the multi-agent system behind the company's decision intelligence product for institutional investors. It consistently outperformed frontier models on the work it was built for, and its text-to-SQL agent benchmarks at top-10-level performance on BIRD. Earlier, at FinOpsly, he took charge of the AI work within his first month and shipped the features that became the product's reason to exist for its enterprise customers.
Saksham combines 3 years of hands-on industry experience and multi-venture leadership across AI infrastructure, protocol engineering, and autonomous systems. Before building Last Lab, he worked at Turing, authoring complex bug-detection, code reasoning, and multimodal training for SOTA models including Gemini and Llama. A relentless builder, he has earned victory across multiple national and international hackathons.
Turn a paper, video, playlist, PDF, or repository into verified code with the compute to run it.
Explore Last Lab beyond learning
Join Discord