GSoC 2027 plan for rishabhsai

The short version

Google DeepMind ran a GSoC org exactly once, in 2025. It was not in 2024 and it is not in the 2026 org list. Its 2025 projects were Gemini API and Gemma developer tooling run by Google AI developer relations, not MuJoCo, Optax, or dm_control work. A separate Gemini CLI org participated in 2026 with two projects, but Google moved individual users off Gemini CLI in June 2026, so its 2027 return is uncertain.

So the plan has two tracks. Track one is Kubeflow as the primary target, with CNCF as the second proposal, because prior contributions decide selection there and the student's Go, Python, and TypeScript fit. Track two is a Google-adjacent watch list (DeepMind, Gemini CLI) that the same Gemini cookbook and SDK contributions serve. The single most important thing to do before December 2026 is to land three or more non-trivial merged PRs in one primary org and become a known name in its chat, because every mentor-side source says orgs do not select people they have not talked to.

Timeline, actual and projected

The 2026 row is the current official timeline, which also sets an org discussion period from Feb 19 to Mar 15, rankings due Apr 21, community bonding May 1 to 24, midterms Jul 6 to 10, final submissions Aug 17 to 24, and an extended-project deadline of Nov 2. The 2024 and 2025 rows come from Google's announcement posts and the Jenkins and Kubeflow pages for those years.

The pattern is stable. Org applications open on a Monday in the second half of January and close about 15 days later, orgs are announced about two weeks after that, the contributor window opens three to four weeks later and lasts 15 days, and acceptance comes about 30 days after the deadline. Google published the 2026 dates in December 2025, so expect the 2027 dates in December 2026.

Rules that matter for this student

Eligibility per the 2026 rules page. You must be 18 or older at registration, eligible to work in the country where you reside, not an org admin or mentor, and either a student or a beginner to open source. You cannot have been accepted to GSoC two or more times before, and Google employees and interns are ineligible. An enrolled undergraduate qualifies through the student clause, so contribution count before applying does not affect eligibility. The FAQ's beginner definition (fewer than 10 issues or PRs, "Regular contributors to an open source project are not beginners") only matters for non-students. If you are on a student visa, check the work-eligibility wording with Penn State's international office before registering.

Proposals. Up to three per contributor, only one can be accepted. Proposals do not have to come from an org's ideas list. If you propose work you already started, you must say so, and that prior work does not count toward evaluations.

Project sizes. Small ~90 hours, medium ~175, large ~350. In 2025, 53.7 percent of accepted projects were large, 41.5 percent medium, 4 percent small, and 18 percent ran an extended schedule.

Stipends for 2026, US residents. $1,500 small, $3,000 medium, $6,000 large, paid 45 percent after the midterm and 55 percent after the final evaluation. Other countries are PPP-adjusted within published bands.

AI use. There is no single program rule. Google's 2026 guidance says "Use AI tools mostly for research and less for code generation" and "Always validate what the AI generates, and if you don't understand it or are not sure, don't use it." The FAQ warns that an AI-written proposal "could result in an automatic rejection by the organization," and the org admin tips tell orgs to ignore low-quality AI proposals. Open Robotics goes further and discourages AI-generated applications and forbids AI-generated code. Read each org's contributor guidance link before writing anything.

What Google DeepMind's GSoC org actually was

Participation. Not in the 2024 archive (195 orgs). Present in 2025 as google-deepmind-sq with 38 archived projects. Absent from the live 2026 org list (183 orgs on 2026-09-11), the 2026 org lookup returns not found, and gsocorganizations.dev lists 2025 as its only year.

Who ran it. The 2025 contact was webpaige@google.com, and the ideas list resolved to a gist by Paige Bailey of Google AI developer relations. Tech tags were python, javascript, typescript, Jax, Gemma.

What the ideas list asked for. A Gemini Workspace in Postman, evals documentation for the Gemini APIs, Gemini integrations in open source agent tools and editor extensions, a batch prediction sample using long context and context caching, example apps, evaluating Gemini on an open benchmark, OSS-Fuzz and OSS-Fuzz-Gen work mentored by Dongge Liu, and Gemma projects (fine-tuning UI, function calling, Hugging Face Spaces demo, benchmarks), sized 175 to 350 hours. The proposal ask was one sentence, that proposals "demonstrate your understanding of the project, your proposed approach, and your relevant skills." The gist also said in bold that applicants must not contact GSoC organizers, DeepMind researchers, engineers, or anyone at Google outside official GSoC channels.

What was accepted. Thirty-eight projects, nearly all Gemini API or Gemma developer tooling, for example a Gemini cookbook migration to JavaScript and TypeScript, Gemini examples in Swift, a Gemini-powered VS Code extension, two batch-prediction and context-caching projects, several benchmarks, Gemma fine-tuning platforms, an Xarray-JAX library, and agent architectures for Concordia. Most mentors were listed publicly as "Google DeepMind." Contributor reports name Paige Bailey and Vaibhav Tulsyan. There were no MuJoCo, MJX, Optax, Penzai, dm_control, or TreeScope projects.

What this means for the plan. The in-flight PRs to google-deepmind/mujoco and google-deepmind/optax will not map to a DeepMind ideas list, because none existed for those repos. They still help, since Optax work shows the JAX literacy the org tagged and both show you can land code in DeepMind-maintained repos, which belongs in the "relevant skills" section of any proposal. But the contributions that mirror its accepted projects are in google-gemini/cookbook, googleapis/js-genai, and Gemma tooling on Hugging Face Transformers. One 2025 contributor's final report shows exactly that mix.

Gemini CLI, the Google org that did show up in 2026

Gemini CLI was a first-time org in 2026 (TypeScript, tags GenAI, MCP, agents), with contributor guidance at the repo's CONTRIBUTING.md. The maintainer's March 2026 update said the January ideas were already implemented by March, so they published a smaller list as GitHub issues and told applicants to reply there, adding that "All other direct contacts will be ignored given the volume of interest." Two projects were accepted and passed, a long-context coding evaluation dataset (medium, mentor Akhilesh Kumar) and behavioral evals and quality work (large, mentor Christian Gunderman).

On May 19, 2026 Google announced Antigravity CLI, built in Go, and said that on June 18, 2026 Gemini CLI would stop serving free, AI Pro, and AI Ultra users, with enterprise licenses unaffected. The repository is still public and active (pushed 2026-09-11, not archived, about 107k stars). Whether the team applies for 2027 is unknown. Finish the parallel gemini-cli PR, since it is the only Google org that participated in 2026, but do not build the plan around it.

Other orgs, ranked for this student

Project counts are accepted projects in 2024, 2025, and 2026 from the GSoC site's archive and live APIs. Realism is a 1 to 10 judgment of a well-prepared applicant's chance.

Also worth a look. Apache Software Foundation (20, 25, 34 projects, JIRA-based ideas, some Go, Python, and TypeScript, mostly Java), ML4SCI (26, 31, 31, Python ML with an evaluation test per project, physics and astronomy), Python Software Foundation (30, 16, 13, bug-fix PR required before applying, Pwndbg is the dev-tooling sub-org), Open Robotics (5, none, 7, the MuJoCo-adjacent choice, but C++ heavy and hostile to AI-assisted work), and The Linux Foundation (19, 21, 18).

Orgs to drop. Chromium was in 2024 and 2025 but not 2026. JAX and Keras ran 4 projects in 2025 only and is absent in 2026. TensorFlow, Mozilla, and OpenMined are absent from all three years. PyTorch, Hugging Face, and Google itself have never been orgs in these lists. LLVM (12, 13, 11) and Julia (19, 13, 12) need C++ or Julia.

What mentors say they select on

The GSoC mentor guide says "Don't even think about selecting a GSoC contributor with whom you've had no contact" and that mentors should establish "an active back-and-forth prior to making a decision." It tells mentors to send an early query and treat a slow reply as a warning sign, to watch community engagement, and not to decide on proposal quality alone.

Google's org admin tips repeat it in capitals, "NEVER select a GSoC contributor based solely on what they wrote in their proposal" and "ALWAYS talk to every GSoC contributor you are considering selecting." Orgs may require a PR as a pre-task and filter out anyone who skipped it.

CNCF's admin posted its 2023 criteria. A meaningful prior contribution, a proposal that shows understanding of the problem, community engagement, and interest from more than one mentor ranked higher. Typo fixes did not count. The Python Software Foundation says "Usually we expect GSoC contributors to fix a bug and have made a pull request" before applying and requires a link to it. Google's 2026 org announcement said "Chatting and contributing early is HUGE for getting accepted," to read each org's contributor guidance link for pre-tasks, and to submit three days early.

The contributor guide adds structure. Make first contact before the application period, introduce yourself, ask real questions, and be humble for a few months. A proposal needs a clear title, synopsis, community benefit, deliverables with milestones, related work, and a biography with evidence of your qualifications. Under-scoped and finished beats ambitious and unfinished. One DeepMind 2025 report lists PRs in the Gemini cookbook and SDK repos as the substance of the project, and another's project pivoted after mentor discussion, which only happens when a conversation exists.

Month by month, September 2026 to May 2027

Odds and failure modes

Program-wide, 2025 had 23,559 proposals from 15,240 applicants and 1,280 accepted, about 5.4 percent of proposals and 8.4 percent of applicants. 2026 had 23,371 proposals from 15,245 applicants and 1,141 accepted, about 4.9 percent. Of 2025's accepted contributors, 43 percent had no prior open source experience and 89 percent were enrolled students, so being new is normal.

DeepMind is a long shot for structural reasons. It skipped 2026 after one year, 38 slots is small against the volume a Google brand attracts, its rules forbid the early mentor relationships every other source says drive selection, and its projects were developer-relations work rather than research library work. Gemini CLI had 2 slots and its product was withdrawn from individual users in June 2026.

Ways this plan fails

• Proposing to orgs you have not talked to.
• Counting typo fixes as contributions.
• Using AI to draft a proposal for an org that forbids it.
• Splitting effort across five orgs so no maintainer knows you.
• Starting in February instead of September.
• Slow replies in April.
• Proposing a large project when a medium one you can finish would rank higher.

Fallback list in order

1. Kubeflow
2. CNCF
3. NumFOCUS (pytorch-ignite or sbi)
4. OpenVINO
5. Apache (Go or TypeScript items)
6. ML4SCI
7. Python Software Foundation (Pwndbg)
8. The Linux Foundation