A real sample dossier

Generated for Dr. Priya Raman, a fictional ML-researcher-turned-founder persona (Dr. Priya Raman is not a real person). The intake answers are invented; the six documents are the production pipeline's unedited output for them.

This sample shows the full, unedited dossier in the format a paying customer receives. Every document carries its attestation line; criteria with no evidence are rated Absent, not inflated. The letter briefs are skeletons with bracketed placeholders, never ghost-written letters.

The dossier itself is written in English, because USCIS requires filings in English. This site is translated so you can read it; the documents you receive are not.

Cover Note to the Reviewing Attorney

Based on evidence provided by Dr. Priya Raman on 2026-10-28. Prepared by an automated evidence-research service, not a law firm; not legal advice; no visa-route recommendation; no USCIS outcome prediction; every claim requires independent verification by the reviewing attorney.

This dossier assembles evidence for the O-1A extraordinary ability route that Dr. Priya Raman reported her attorney is preparing. It is research and assembly support of the kind a paralegal might provide. It is not a petition, not legal advice, not prepared by a law firm, not a guarantee of any USCIS outcome, and not a recommendation of any visa route. The choice of route, and every judgment about eligibility and strategy, remains with you.

Source of facts. Every factual statement comes from Dr. Raman's intake answers of 2026-10-28, and each one cites the intake question it came from. The service has not independently confirmed any item. Each claim carries a verification hook, meaning the document or record that would substantiate it. Examples include her Google Scholar profile for the reported ~5,800 citations and h-index of 16, arXiv:2210.11416 for the FLAN paper, the ACL 2022 workshop records for the best-paper honorable mention, and the NeurIPS and ICML reviewer records. Please treat each hook as a step still to be completed.

Mapping and ratings. The evidence is mapped only to the eight criteria at 8 CFR 214.2(o)(3)(iii)(B)(1)–(8). Each criterion has a strength rating of Strong, Moderate, Thin, or Absent. These ratings describe only how much evidence the intake supplied. They are evidence-availability ratings. They are not eligibility judgments and not outcome predictions.

Gaps. The applicant marked some items as not yet in place. These appear only in the gap analysis, where they are described as gaps rather than as evidence.

Recommendation-letter briefs. The briefs for Prof. Andreas Keller, Dr. Sofia Marchetti, and James Wu are skeletons. They contain factual anchors from the intake and bracketed placeholders for each recommender's own knowledge and assessment. No recommender opinion has been drafted on their behalf.

Contact. The applicant's delivery contact is priya.raman@example.com.

Applicant Profile Summary

Purpose of This Section

This section is a factual, third-person summary of the applicant's professional background, drawn only from the intake answers. It is not legal advice and was not prepared by a law firm. It is not a guarantee of any USCIS outcome and not a recommendation of any visa route. The petition route shown below (O-1A) is the route the applicant reports her attorney is preparing. It is not a selection made by this service.

Identification

Name: Dr. Priya Raman
Field of work (as stated by the applicant): Machine-learning infrastructure
Current role: Co-founder & CTO at an 8-person ML infrastructure startup (YC W24)
Highest degree: PhD in Computer Science, ETH Zürich, 2019
Petition route being prepared by counsel (per intake): O-1A

Professional Background

Dr. Raman completed doctoral research at ETH Zürich from 2015 to 2019. Her PhD addressed distributed training of large neural networks. The intake names Prof. Andreas Keller as her thesis advisor.

From 2019 to 2023, she worked as a Research Scientist at Google DeepMind in Zürich. The intake describes Google DeepMind as a leading AI research organization and states that she worked on the FLAN project. She is a co-author of "Scaling Instruction-Finetuned Language Models" (FLAN), arXiv:2210.11416, 2022. The intake reports that this paper has been cited more than 4,000 times.

Since 2023, she has been Co-founder & CTO of an 8-person ML infrastructure startup. The startup took part in Y Combinator's Winter 2024 batch and raised a $4.5M seed round. The intake states that she owns the core training-infrastructure architecture on which the company's product is built.

Employment History (most recent first, as reported)

Period	Role	Organization
2023–present	Co-founder & CTO	ML infrastructure startup (YC W24)
2019–2023	Research Scientist	Google DeepMind (Zürich)
2015–2019	Doctoral researcher	ETH Zürich

Research and Technical Activity (summary only)

The intake reports the following:

Publications: Authorship of the FLAN paper and 7 further peer-reviewed papers at NeurIPS, ICML, and ACL, published 2019–2024.
Citations: About 5,800 total citations on Google Scholar and an h-index of 16.
Open-source work: Core contributor to a distributed-training library with 12k GitHub stars, which is used in production by research labs. She also maintains two personal open-source projects with about 1.5k stars combined.

These items are listed here for context only. They are mapped to specific criteria in other sections of this dossier.

Proposed Endeavor (in the applicant's own terms)

The applicant describes her intended U.S. work as building distributed-training infrastructure that lets small research teams train large models without hyperscaler budgets. The company serves university labs and early-stage startups. She describes this work as a continuation of her published research on distributed training of large neural networks.

Verification Notes for the Reviewing Attorney

The reviewing attorney should independently verify each of the following:

Startup records: The company's name, its YC W24 participation, and its seed funding amount. The intake does not name the company.
Citation metrics: Current Google Scholar figures.
Publication list: The specific titles and venues of the 7 further papers, which the intake does not list individually.
Open-source metrics: The identity of the distributed-training library and its current star count.

Criteria Mapping — Dr. Priya Raman (O-1A extraordinary ability)

Scope and limits of this document. This table maps intake answers to the O-1A evidentiary criteria at 8 CFR 214.2(o)(3)(iii)(B) for the reviewing attorney. It is not legal advice. It is not prepared by a law firm. It is not a guarantee of any USCIS outcome. It is not a recommendation of any visa route. The intake states that the attorney is preparing an O-1A petition. This mapping follows that choice and does not evaluate it.

About the strength ratings. Each rating (Strong / Moderate / Thin / Absent) describes only how much documentable evidence the intake supplied for that criterion. A rating is not an eligibility judgment. It is not a prediction of how USCIS would weigh the evidence. Whether any criterion is met is a legal determination for the reviewing attorney.

Awards or prizes for excellence (8 CFR 214.2(o)(3)(iii)(B)(1))

Evidence from intake:

Best-paper honorable mention at an ACL 2022 workshop. The intake does not name the specific workshop or paper.

Strength: Thin. This is an evidence-availability rating only. The intake supplies one item. It is a workshop-level honorable mention rather than a main-conference award. The intake gives no information on the size of the selection pool or the extent of the award's recognition.

Verification hooks:

Identify the specific ACL 2022 workshop and the paper that received the honorable mention.
Obtain the award notice, certificate, or official announcement from the workshop organizers.
Archived workshop webpage or proceedings listing the honorable mention (ACL Anthology URL, if available).
Any documentation of selection criteria, number of submissions, and the award committee's composition.

Membership in associations requiring outstanding achievement (8 CFR 214.2(o)(3)(iii)(B)(2))

Evidence from intake: None provided in intake.

Strength: Absent. This is an evidence-availability rating only. See the Gap Analysis section below.

Verification hooks: Not applicable at this stage. See the Gap Analysis section below.

Published material about the applicant (8 CFR 214.2(o)(3)(iii)(B)(3))

Evidence from intake:

Quoted in TechCrunch coverage of the company's seed round (2024).
Invited guest on a well-known ML podcast (2024). The intake reports about 50,000 listeners and does not name the podcast.

Strength: Thin. This is an evidence-availability rating only.

The TechCrunch item is described as a quote within coverage of the company's funding round. Whether the article is about Dr. Raman herself is not established by the intake.
The podcast appearance is an interview rather than published material written about her.
The listener figure is self-reported.

Verification hooks:

Full copy of the TechCrunch article, with URL, publication date, and author byline. Confirm how much of the article concerns Dr. Raman versus the company.
Evidence of TechCrunch's circulation or readership, so the attorney can assess its status as major media.
The podcast's name, episode title, date, and URL or transcript.
Independent source for the ~50k listener figure, such as host-provided analytics or published audience data.

Judging the work of others (8 CFR 214.2(o)(3)(iii)(B)(4))

Evidence from intake:

Reviewer for NeurIPS in 2022, 2023, and 2024.
Reviewer for ICML in 2023 and 2024.
Self-described "area-chair-adjacent load of 15+ papers per year."

Strength: Strong. This is an evidence-availability rating only. The intake names specific venues and years across multiple review cycles, and conference reviewing records are generally documentable. The "area-chair-adjacent" description and the 15+ papers per year figure are self-characterizations that need documentation.

Verification hooks:

Reviewer invitation emails and acceptance confirmations from NeurIPS (2022–2024) and ICML (2023–2024).
OpenReview or CMT reviewer-console records showing assigned papers per cycle, which would support the 15+ papers per year figure.
Any published reviewer lists or acknowledgments from those conferences.
Documentation of the actual role held (reviewer vs. any senior or area-chair designation). The intake indicates reviewer, not area chair.

Original contributions of major significance (8 CFR 214.2(o)(3)(iii)(B)(5))

Evidence from intake:

Co-author of "Scaling Instruction-Finetuned Language Models" (FLAN), arXiv:2210.11416, 2022. The paper is cited 4,000+ times per the intake.
Core contributor to a distributed-training library with 12k GitHub stars that is used in production by research labs. The intake does not name the library.
Maintains two personal open-source projects with about 1,500 GitHub stars combined.
The current company builds distributed-training infrastructure, described as continuing her PhD research on distributed training of large neural networks.
James Wu is CEO of a partner company that uses the startup's training infrastructure in production.

Strength: Moderate. This is an evidence-availability rating only.

Citation and adoption figures are concrete.
The intake does not specify Dr. Raman's individual contribution within the FLAN author list or within the shared library.
"Major significance" typically depends on documented field impact and expert letters, which are not yet in hand.

Verification hooks:

Google Scholar page for the FLAN paper showing the current citation count, as a dated screenshot or PDF.
Author contribution statement for FLAN, if one exists, or a statement from co-author Dr. Sofia Marchetti describing Dr. Raman's specific role.
The library's name, GitHub URL, and star count (dated capture). Commit history or contributor graph showing Dr. Raman's "core contributor" status.
Evidence of production use by research labs, such as user lists, citing papers, or adoption statements.
GitHub URLs and dated star counts for the two personal projects.
Written confirmation from James Wu's company of production use of the startup's infrastructure.

Authorship of scholarly articles (8 CFR 214.2(o)(3)(iii)(B)(6))

Evidence from intake:

"Scaling Instruction-Finetuned Language Models" (FLAN), arXiv:2210.11416, 2022.
Seven further peer-reviewed papers at NeurIPS, ICML, and ACL, 2019–2024. The intake does not list individual titles.
Google Scholar: about 5,800 total citations, h-index 16.

Strength: Strong. This is an evidence-availability rating only. The intake identifies eight papers, top ML venues, and quantified citation metrics. Individual titles for the seven additional papers still need to be supplied. The FLAN paper is cited by arXiv identifier, so its peer-reviewed venue, if any, should be confirmed.

Verification hooks:

Full bibliographic list of all eight papers, with titles, co-authors, venue, year, and DOI or proceedings URL.
Copies of each paper's first page showing Dr. Raman's authorship.
Dated Google Scholar profile capture showing ~5,800 citations and h-index 16.
Confirmation of the FLAN paper's publication venue beyond arXiv.
Evidence that NeurIPS, ICML, and ACL are professional or major venues in the field, such as acceptance rates or venue descriptions.

Critical role for distinguished organizations (8 CFR 214.2(o)(3)(iii)(B)(7))

Evidence from intake:

The startup. Co-founder & CTO, 2023–present, of an 8-person ML infrastructure startup. The startup was in Y Combinator W24 and raised a $4.5M seed round. She owns the core training-infrastructure architecture the product is built on.
Google DeepMind. Research Scientist at Google DeepMind (Zürich), 2019–2023, on the FLAN project.

Strength: Moderate. This is an evidence-availability rating only.

The roles and organizations are named. The CTO role is described concretely.
The startup's distinguished reputation would need documentation; YC participation and seed funding are the only intake indicators.
The intake does not describe the specific critical function she performed at DeepMind beyond working on the FLAN project.

Verification hooks:

Startup incorporation and role. Incorporation documents and a cap table or founder agreement showing Dr. Raman's co-founder & CTO role.
YC participation. YC W24 batch listing or YC company page (URL).
Seed round. Seed-round documentation, such as a term sheet or investor confirmation, and the TechCrunch article reporting the $4.5M round.
Architecture ownership. A letter from the CEO or co-founder, or technical documentation, describing her ownership of the core training-infrastructure architecture.
DeepMind employment. Employment verification letter for 2019–2023 stating title and responsibilities.
FLAN role. Statement from Dr. Sofia Marchetti, former research lead, on the role Dr. Raman held on FLAN.
DeepMind reputation. Evidence of Google DeepMind's reputation in the field.

High salary or remuneration (8 CFR 214.2(o)(3)(iii)(B)(8))

Evidence from intake: None provided in intake.

Strength: Absent. This is an evidence-availability rating only. See the Gap Analysis section below.

Verification hooks: Not applicable at this stage. See the Gap Analysis section below.

Gap Analysis

This section covers items the intake marked as NOT YET IN PLACE. They are listed here as gaps only. None of them is existing evidence.

Patents. No evidence provided. The intake answer was "not yet." No patent evidence is available for the original-contributions criterion. Any reliance on patents would need filed or granted patent documents, which do not currently exist per the intake.
Memberships (criterion 2). No evidence provided. The intake answer was "not yet." To address this criterion, the record would need two things:
- documentation of admission to an association that requires outstanding achievement, as judged by recognized experts;
- that association's published admission criteria.
Salary or remuneration (criterion 8). No evidence provided. The intake answer was "not yet." To address this criterion, the record would need two things:
- compensation documentation, such as pay records, tax records, an offer letter, or equity valuation;
- comparative compensation data for ML infrastructure or research roles.
Recommendation letters. Three potential recommenders are named in the intake. No letters exist yet.
- Prof. Andreas Keller, PhD thesis advisor at ETH Zürich.
- Dr. Sofia Marchetti, former research lead at Google DeepMind and FLAN co-author.
- James Wu, CEO of a partner company using the startup's infrastructure in production.
Any assessment of significance or impact would need to come from the recommenders in their own words.
Unnamed specifics. The intake does not name the following, and each would need to be supplied:
- the ACL 2022 workshop;
- the titles of the seven additional papers;
- the 12k-star library;
- the podcast.

Honest Gap Analysis

This section is not legal advice, not a guarantee of any USCIS outcome, and not a recommendation of any visa route. It lists the criteria where the intake supplied little or no evidence and describes the kinds of documents that could fill those gaps. Ratings (Thin / Absent) are evidence-availability ratings only. They describe how much evidence the intake contained. They are not eligibility judgments or outcome predictions.

Criteria Rated Absent

Membership in associations requiring outstanding achievement (8 CFR 214.2(o)(3)(iii)(B)(2)): Absent (evidence availability only)

Intake status: The memberships answer was marked NOT YET IN PLACE. No evidence was provided.
Evidence that would close the gap:
- Documentation of membership in an association whose admission requires outstanding achievement, as judged by recognized experts.
- The association's published admission criteria.
- Evidence of who evaluates candidates for admission.

High salary or remuneration (8 CFR 214.2(o)(3)(iii)(B)(8)): Absent (evidence availability only)

Intake status: The salary answer was marked NOT YET IN PLACE. No evidence was provided.
Evidence that would close the gap:
- Documentation of current compensation as Co-founder & CTO, such as salary, equity, or board-approved compensation records.
- Comparative wage data for machine-learning infrastructure roles from an objective survey or government source.
- If relevant to the attorney's approach, prior compensation records from Google DeepMind (Zürich).

Criteria Rated Thin

Awards or prizes for excellence (8 CFR 214.2(o)(3)(iii)(B)(1)): Thin (evidence availability only)

Intake status: One item was provided: a best-paper honorable mention at an ACL 2022 workshop.
Evidence that would close the gap:
- Documentation of the award's selection process, number of submissions, and selecting committee.
- Evidence of the award's national or international recognition.
- Any additional prizes or awards for excellence in the field, if they exist.

Published material about the applicant (8 CFR 214.2(o)(3)(iii)(B)(3)): Thin (evidence availability only)

Intake status: Two items were provided:
- A quote in TechCrunch coverage of the company's seed round (2024).
- An invited guest appearance on an ML podcast (2024) with about 50k listeners.
Limitations of the current evidence:
- The TechCrunch item is described as coverage of the company's funding, with the applicant quoted. It is not described as material about the applicant's work.
- The intake does not name the podcast.
Evidence that would close the gap:
- Full copies of published material primarily about Dr. Raman or her work, with title, date, author, and publication.
- Circulation or audience data showing that each outlet is a professional publication, major trade publication, or other major media.
- The podcast's name, an episode transcript, and independent audience figures.

Other Gaps From the Intake

Patents

Intake status: Marked NOT YET IN PLACE. No evidence was provided.
Patents are not a standalone O-1A criterion.
If any patent filings exist or are later made, they could serve as supporting documentation for original contributions of major significance (8 CFR 214.2(o)(3)(iii)(B)(5)).

Supporting Gaps in Criteria Not Rated Thin or Absent

These are gaps within criteria that have more evidence. They are noted for completeness.

Original contributions of major significance (B)(5): The intake gives citation figures (~5,800 total, h-index 16, FLAN 4,000+) and GitHub stars (12k library, ~1.5k personal projects). It does not name the distributed-training library or the research labs using it in production. Useful documents would include:
- Named adopters.
- Usage documentation.
- Independent statements from those adopters.
Judging the work of others (B)(4): The intake lists NeurIPS 2022–2024 and ICML 2023–2024 reviewing. Useful documents would include:
- Invitation letters.
- Confirmation emails or reviewer acknowledgments.
- Support for the "15+ papers per year" figure.
Critical role for distinguished organizations (B)(7): Useful documents would include:
- Organizational charts.
- Letters describing the applicant's role at Google DeepMind on the FLAN project.
- Letters describing her ownership of the startup's core training-infrastructure architecture.
- Evidence of each organization's reputation.

The reviewing attorney should independently decide whether and how any of these gaps matter.

Evidence Register — Dr. Priya Raman

Scope note: This register lists discrete evidence items drawn only from Dr. Raman's intake answers. It is an evidence inventory for attorney review. It is not legal advice, not prepared by a law firm, not a guarantee of any USCIS outcome, and not a recommendation of any visa route. Criterion mappings show where each item may be relevant under the O-1A criteria at 8 CFR 214.2(o)(3)(iii)(B). They are not eligibility determinations. Items the intake marked as not yet in place are excluded here and belong only in the gap analysis.

1. Co-authorship of "Scaling Instruction-Finetuned Language Models" (FLAN), arXiv:2210.11416, 2022 — supports Authorship of scholarly articles (B)(6); Original contributions of major significance (B)(5) — verify via:

the full paper with the author list showing Dr. Raman's name;
confirmation of any later peer-reviewed or journal version, since arXiv is a preprint server and its status as "professional journals or other major media" needs attorney assessment;
a statement from a co-author (e.g., Dr. Sofia Marchetti) describing Dr. Raman's specific contribution.

2. FLAN paper cited 4,000+ times — supports Original contributions of major significance (B)(5); Authorship of scholarly articles (B)(6) — verify via:

a dated Google Scholar screenshot or export for arXiv:2210.11416;
a sample of citing works showing how FLAN is built on or used;
field-normalized citation context, if the attorney finds it useful.

3. Google Scholar profile: ~5,800 total citations, h-index 16 — supports Original contributions of major significance (B)(5); Authorship of scholarly articles (B)(6) — verify via:

a dated, full-page capture of the Google Scholar profile;
a check that every listed paper is correctly attributed to Dr. Raman;
the share of total citations coming from the FLAN paper versus her other work.

4. Seven further peer-reviewed papers at NeurIPS, ICML, and ACL, 2019–2024 — supports Authorship of scholarly articles (B)(6) — verify via:

full citations for each of the seven papers (titles, venues, years, author positions), which the intake does not itemize;
copies of each paper;
proceedings listings confirming that each paper was peer-reviewed and accepted, rather than a workshop or preprint version.

5. Best-paper honorable mention, ACL 2022 workshop — supports Awards or prizes for excellence (B)(1) — verify via:

the name of the specific workshop;
the award certificate or an official announcement;
the selection criteria and the number of submissions;
evidence on whether the honor is nationally or internationally recognized, since workshop-level recognition needs attorney assessment against the criterion's standard.

6. Reviewer for NeurIPS 2022–2024 and ICML 2023–2024, reviewing 15+ papers per year — supports Judging the work of others (B)(4) — verify via:

reviewer invitation emails or acknowledgment letters for each year;
OpenReview or CMT records showing assignment counts;
documentation behind the "area-chair-adjacent" load described in the intake. This should be confirmed as the applicant's own description, not a formal area-chair appointment.

7. Core contributor to a distributed-training library with 12k GitHub stars, used in production by research labs — supports Original contributions of major significance (B)(5) — verify via:

the library's name and repository URL, which the intake does not specify;
commit history and contributor statistics showing Dr. Raman's role;
a dated capture of the star count;
documentation or statements identifying the research labs that use it in production.

8. Two personal open-source projects with ~1.5k GitHub stars combined — supports Original contributions of major significance (B)(5) — verify via:

the project names and repository URLs;
dated star counts;
any evidence of downstream adoption. The attorney should assess whether this item adds weight beyond item 7.

9. Quoted in TechCrunch coverage of the company's 2024 seed round — supports Published material about the applicant (B)(3) — verify via:

the article URL, headline, date, and author;
the full text, to judge whether the article is about Dr. Raman or primarily about the company or the funding round;
TechCrunch circulation or readership data to support "major media" status.

10. Invited guest on an ML podcast (2024), ~50k listeners as reported in the intake — supports Published material about the applicant (B)(3) — verify via:

the podcast name, episode title, date, and link;
independent audience data for the ~50k figure;
a transcript;
attorney assessment of whether a podcast appearance qualifies as published material about the applicant in major media.

11. Co-founder & CTO (2023–present) of an 8-person ML infrastructure startup, YC W24, $4.5M seed round; owns the core training-infrastructure architecture the product is built on — supports Critical role for distinguished organizations (B)(7) — verify via:

incorporation and founding documents;
Y Combinator W24 batch confirmation;
seed-round documentation such as a term sheet or investor announcement;
architecture or technical documentation showing Dr. Raman's ownership;
evidence of the company's reputation, such as the TechCrunch coverage in item 9;
a statement from James Wu (CEO of a partner company using the infrastructure in production).

12. Research Scientist, Google DeepMind (Zürich), 2019–2023, working on the FLAN project — supports Critical role for distinguished organizations (B)(7); provides context for (B)(5) — verify via:

an employment verification letter with title and dates;
documentation of her responsibilities on FLAN;
a statement from Dr. Sofia Marchetti (former research lead, FLAN co-author) on the role. Whether the role was "critical or essential" must come from the recommender's own knowledge;
standard reputational evidence for the organization.

13. PhD in Computer Science, ETH Zürich, 2019, on distributed training of large neural networks (doctoral researcher 2015–2019) — supports background and context for Original contributions (B)(5); continuity with the proposed endeavor — verify via:

the diploma and transcript;
the thesis title and abstract;
a statement from Prof. Andreas Keller (thesis advisor) on the research;
a documented link between the thesis work and the current distributed-training infrastructure described in the proposed endeavor.

Excluded from this register: the intake marked patents, professional memberships, and salary data as not yet in place. They are addressed only in the gap analysis.

Recommendation-Letter Briefs — Dr. Priya Raman

How to use these briefs

These briefs are skeletons, not letters. Each one lists facts from Dr. Raman's intake answers that the named recommender could plausibly address, followed by bracketed placeholders for content only the recommender can supply.

The factual anchors come only from the intake. The applicant and the reviewing attorney should confirm each anchor before it is shared with a recommender.
The bracketed sections are intentionally left unwritten. The assessments, examples and comparisons must come from the recommender's own knowledge and in the recommender's own words.
Each anchor is tagged with the O-1A criterion it relates to under 8 CFR 214.2(o)(3)(iii)(B). These tags are organizational only. They are not judgments of whether any criterion is met.
Whether any letter addresses a particular criterion, and how, is for the reviewing attorney to decide.
This document is not legal advice, not prepared by a law firm, not a guarantee of any USCIS outcome, and not a recommendation of any visa route.

Brief 1 — Prof. Andreas Keller

Relationship to applicant: PhD thesis advisor, ETH Zürich (per intake).

Factual anchors the letter can cover:

Doctoral training
- Dr. Raman was a doctoral researcher at ETH Zürich from 2015 to 2019.
- She received a PhD in Computer Science from ETH Zürich in 2019.
Thesis topic
- Her doctoral research was on distributed training of large neural networks.
- Relevant to: Original contributions of major significance, (B)(5).
Continuity with proposed U.S. work
- The intake states that her current work is building distributed-training infrastructure for small research teams.
- It describes this as a continuation of her published research on distributed training of large neural networks.
- Relevant to: Original contributions, (B)(5).
Scholarly output (publication years overlap with the doctoral period)
- She has 7 peer-reviewed papers at NeurIPS, ICML and ACL dated 2019–2024, in addition to the FLAN paper.
- The intake does not say which, if any, came from the doctoral work. The applicant should confirm this before the anchor is used.
- Relevant to: Authorship of scholarly articles, (B)(6).
Later trajectory, which the advisor may or may not have followed
- Research Scientist at Google DeepMind (Zürich), 2019–2023.
- Co-founder & CTO of an 8-person ML infrastructure startup, 2023–present.

[For the recommender to complete]:

[Prof. Keller to describe in his own words his role, his position at ETH Zürich, and how long he supervised Dr. Raman.]
[Prof. Keller to describe the specific problem the dissertation addressed and what, in his own view, was original about her approach.]
[Prof. Keller to identify any thesis results, papers or methods he personally knows were adopted, extended or cited by others, with specifics he can verify.]
[Prof. Keller to state, only if he chooses to, any comparison between Dr. Raman and other doctoral researchers he has supervised or observed, using his own basis for comparison.]
[Prof. Keller to add any first-hand examples of her independent research contributions during 2015–2019.]
[Prof. Keller to state his own credentials, so the reader can understand the basis for his views.]

Brief 2 — Dr. Sofia Marchetti

Relationship to applicant: Former research lead at Google DeepMind; FLAN co-author (per intake).

Factual anchors the letter can cover:

Employment at Google DeepMind
- Dr. Raman was a Research Scientist at Google DeepMind (Zürich) from 2019 to 2023.
- She worked on the FLAN project.
- Relevant to: Critical role for distinguished organizations, (B)(7).
The FLAN paper
- Full reference: Scaling Instruction-Finetuned Language Models (FLAN), arXiv:2210.11416, 2022.
- Dr. Raman is a co-author.
- Relevant to: Authorship, (B)(6); Original contributions, (B)(5).
Citation metrics (per intake; to be independently verified)
- The FLAN paper is cited 4,000+ times.
- Dr. Raman's Google Scholar profile shows about 5,800 total citations and an h-index of 16.
- Relevant to: Original contributions, (B)(5).
Peer reviewing, if Dr. Marchetti has direct knowledge of it
- Reviewer for NeurIPS 2022–2024 and ICML 2023–2024.
- Review load of 15+ papers per year.
- Relevant to: Judging the work of others, (B)(4).
Award, only if Dr. Marchetti has direct knowledge of it
- Best-paper honorable mention at an ACL 2022 workshop.
- The intake does not name the paper or say whether it relates to DeepMind work.
- Relevant to: Awards, (B)(1).

[For the recommender to complete]:

[Dr. Marchetti to describe her own role and seniority at Google DeepMind, and her working relationship with Dr. Raman, including dates and team structure.]
[Dr. Marchetti to describe, from first-hand knowledge, which parts of the FLAN work Dr. Raman was responsible for.]
[Dr. Marchetti to state in her own words whether and why she regards Dr. Raman's role on the FLAN project as critical or essential, citing specific examples she witnessed.]
[Dr. Marchetti to describe, if she chooses, any known downstream uses or influence of the FLAN work that she can personally speak to.]
[Dr. Marchetti to give any comparison she is willing to make between Dr. Raman and other research scientists she has worked with, on her own stated basis.]
[Dr. Marchetti to state her own credentials and publication record, so the reader can understand the basis for her views.]

Brief 3 — James Wu

Relationship to applicant: CEO of a partner company that uses the startup's training infrastructure in production (per intake). The intake does not name the partner company; the applicant should supply the name.

Factual anchors the letter can cover:

Production use of the startup's product
- Mr. Wu's company uses the startup's training infrastructure in production.
- Relevant to: Original contributions, (B)(5).
Dr. Raman's role at the startup
- She is Co-founder & CTO, 2023–present.
- Per intake, she owns the core training-infrastructure architecture the product is built on.
- Relevant to: Critical role, (B)(7).
Company facts (per intake)
- The startup is 8 people.
- It is in Y Combinator's W24 batch.
- It raised a $4.5M seed round.
- Relevant to: Critical role, (B)(7).
Product purpose (per intake)
- The infrastructure lets small research teams train large models without hyperscaler budgets.
- The company serves university labs and early-stage startups.
Press coverage, if Mr. Wu is aware of it
- Dr. Raman was quoted in TechCrunch coverage of the seed round (2024).
- Relevant to: Published material, (B)(3).
- The reviewing attorney should assess whether this coverage is about the applicant.

[For the recommender to complete]:

[Mr. Wu to name his company, describe its work, and state his role.]
[Mr. Wu to describe how and since when his company has used the startup's infrastructure, and what it is used for, with concrete details he can verify.]
[Mr. Wu to describe in his own words any effect the infrastructure has had on his company's work, such as cost, time or capability. Only figures he can document should be included.]
[Mr. Wu to describe his direct interactions with Dr. Raman, and any technical decisions or contributions of hers he observed first-hand.]
[Mr. Wu to give his own assessment, if he chooses, of how the infrastructure compares with alternatives his company considered.]
[Mr. Wu to state his professional background, so the reader can understand the basis for his views.]

Reminder: this document is not legal advice, not prepared by a law firm, not a guarantee of any USCIS outcome, and not a recommendation of any visa route. All anchors require verification by the reviewing attorney, and every opinion in a final letter must be the recommender's own.

Build my dossier - $1499 →Free criteria pre-check