Collaboration · Seminars

A methods discussion, not a keynote

For lab seminars, reading groups and methods workshops, I can present a piece of my research and spend most of the time on how it was done: what was measured, what could have gone wrong and what remains open.

The situation

Can we arrange a focused methods discussion or research seminar rather than a commercial keynote?

What you leave with

A research abstract, prerequisite reading and discussion questions for an agreed topic, followed by a seminar weighted towards discussion.

For: Seminar organiser; principal investigator

Topics with research behind them

Each topic below has at least one paper the audience can read beforehand. Status matters, so it is stated.

How reproducible are evaluation conclusions? A self-audit of an LLM evaluation: eight open model variants, 293 persisted raw outputs, and a cluster bootstrap showing that a small-sample ranking identifies the worst model reliably but not the best. Preprint. (paper)
When “correct” is not correct: testing generated GPU kernels. Why fixed-shape allclose checks certify buggy kernels, and what seeded fuzzing against an fp64 reference catches instead. Preprint, with a public 26-op corpus. (paper)
Auditing RAG conflict detectors by input regime. Why the trade-off between an NLI cross-encoder and an LLM judge depends on the input distribution. Manuscript submitted to a workshop. (paper)
Coordination before the pull request. How concurrent coding agents claim and collide over work, studied through a git-native coordination log; duplicate work fell from 78% to 0% in that setting. Preprint. The log, grite, is a Neul Labs project, and I founded Neul Labs; that is stated in the session. (paper)
Federated learning with imbalanced data. Fed-Focal Loss, a workshop paper at the IJCAI 2020 federated learning workshop. (paper)

Decentralised-systems topics — DePIN protocol design, cross-rollup composability, MEV mitigation — draw on earlier preprints listed under publications.

How a seminar differs from a talk

A keynote is a finished performance for a broad audience. A seminar here is closer to a working session: the presentation is short, the reading is done in advance, and the time goes on the parts of the method that could be wrong — sample sizes, baselines, tolerances, what a check can and cannot establish. I would rather spend an hour on one weak assumption than cover five results. For the same reason, I prefer sessions where the host group also brings a methods question of its own.

What a session package looks like

Illustrative example — not a client engagement. Abstract: “Most LLM leaderboards report one number per model. This session re-analyses a small evaluation with a cluster bootstrap and asks which conclusions survive.” Reading: the reproducibility self-audit, sections on the bootstrap and run aggregation. Discussion questions: What sample size would your group need before claiming a best model? Which of your current reported rankings would you be least comfortable re-running? When is a stable worst model still a useful result?

Seminars are where many collaborations start. If one does, it moves to an explicit industry–academic collaboration or consortium contribution with its own terms.

Commission this if

You run a lab seminar, reading group or doctoral methods session and want a speaker who will discuss method in detail.
Your group works on evaluation, benchmark design, kernel testing, federated learning or decentralised protocols and wants an outside view of its own methods.
You want an exchange that might lead to a collaboration, without committing to one.

Not the right fit if

You are booking a conference keynote, panel or corporate event. That is a separate, bookable speaking product.
You want a product or vendor presentation.
You need training delivered to a team as a paid service.

What you receive

Research abstract. A short abstract for the agreed topic, written for your audience's background.
Prerequisite reading. One or two papers, usually my preprints plus a reference from the wider literature, with the sections that matter marked.
Discussion questions. Open methodological questions for the session, including ones where my own work is weakest.
Materials afterwards. Slides and pointers to code and data, where the host and I agree to share them.

What the study needs from you

An agreed topic. Chosen from my published research areas, or a methods question your group is working on.
An audience description. Who attends, what background they have, and how long the session is.
Format. Remote or in person, and whether the session is recorded.

How the work runs

1
Invitation. Send the topic, audience and preferred dates.
2
Agree the topic. We settle the focus and I disclose any relevant interests, such as work connected to Neul Labs.
3
Prepare. I send the abstract, reading and discussion questions in advance.
4
Seminar. A short presentation, then most of the time on discussion.
5
Follow-up. Materials are shared as agreed. If a shared research question emerges, we can discuss collaboration separately.

Limits and unfavourable results

Availability depends on research commitments; not every invitation can be accepted.
A seminar presents my preprints and ongoing work as they stand. Preprints are not peer-reviewed, and the discussion treats them accordingly.
A seminar does not imply an endorsement of the host's work, or the reverse.

Engagement terms

Model: Research exchange by invitation: a seminar or methods discussion on an agreed topic.
Who does the work: Dipankar Sarkar presents and leads the discussion personally.
Commercial basis: A non-commercial research exchange, not a paid speaking engagement. Format and any logistics are agreed per invitation.
Availability: Checked per enquiry.

Conflicts are checked before scoping. See publication, IP and independence.

Evidence behind this page

How Reproducible Are Evaluation Conclusions? A Self-Audit of LLM-Inferred Prompt Structure
2026 · Preprint (not peer reviewed)
The Correctness Illusion in LLM-Generated GPU Kernels
2026 · Preprint (not peer reviewed) · cited by 6 (Google Scholar, October 2026)
An Input-Regime Audit of Conflict Detection for Retrieval-Augmented Generation
2026 · Manuscript under submission
Before the Pull Request: Mining Multi-Agent Coordination
2026 · Preprint (not peer reviewed)
Fed-Focal Loss for imbalanced data classification in Federated Learning
2020 · Peer-reviewed workshop paper · cited by 96 (Google Scholar, October 2026)

Questions

Can we invite Dipankar Sarkar to give a research seminar?

Yes, for lab seminars, reading groups and methods workshops on topics close to my published research. Send the topic, audience and format; availability is checked per invitation.

What topics are available for an AI evaluation methods seminar?

Reproducibility of evaluation conclusions, correctness checking of LLM-generated GPU kernels, conflict detection in retrieval-augmented generation, coordination between concurrent coding agents, and federated learning on imbalanced data. Other topics can be agreed if they connect to my work.

Do you speak at distributed systems or decentralised systems seminars?

Yes, where the discussion is about research: federated learning, DePIN mechanism design, composability and MEV are areas with published preprints or workshop papers behind them.

Is this the same as booking you for a keynote?

No. A seminar is a non-commercial research exchange centred on methods and discussion. Conference keynotes and commercial event talks are booked through dipankar.org.

Send a research question

Send the topic or methods question, the audience and the format you have in mind. A few lines are enough.

Send a research question Draft a brief first

Last reviewed 2026-10-07.