Collaboration · Consortium work package
A work package with edges you can write down
If you are assembling a consortium and need a hands-on AI systems researcher to own a technical work package, I can scope one with explicit dependencies, deliverables and capacity. What I cannot supply is institutional standing: I am an independent researcher, and whether I can take part in a funded call depends on that call and its lead partner.
The situation
What precise work package could Dipankar contribute to a consortium without implying funding eligibility or institutional authority?
What you leave with
A bounded technical contribution statement: objectives, tasks, dependencies on other partners, deliverables and a capacity statement, with eligibility left explicitly to the call rules and the lead partner.
For: Consortium lead; principal investigator
What I can credibly own
A work package I can own has a technical output that other partners depend on, and that one person can build and run. Three shapes fit, each tied to published work (preprints and a submitted manuscript, not funded consortium projects):
- Evaluation infrastructure and benchmark design. A corpus, a harness and a reporting protocol other work packages run against. The gpuemu corpus has 26 ops with seeded bugs and controls; in Test-Input Generation for Tensor Programs it was used to compare seven input-generation strategies on both bug recall and control false-positive rate. The Correctness Illusion used it to show that fixed-shape checks certify buggy kernels.
- An audit of a component across input regimes. Taking a component the consortium relies on and testing where its trade-offs change, as in the RAG conflict-detection audit.
- A measurement substrate for multi-agent experiments. Logging coordination so that failure modes can be mined afterwards, as in Before the Pull Request, which uses grite. grite is a Neul Labs project and I founded Neul Labs; a consortium relying on it should treat that as a disclosed interest.
What I cannot supply
Institutional standing. I am an independent researcher with no institutional signing authority, so I cannot sign agreements, hold funds for a university, coordinate the consortium or make declarations on anyone’s behalf. Whether I can take part in a funded call at all depends on the call and the lead partner; some calls have a route for a contributor like me and some do not. That question is settled first, by the lead partner, before any scoping time is spent.
Anatomy of a bounded contribution
A contribution is bounded when an outsider could read it and say whether it was delivered. That means each task names its inputs, its output artefact, its owner and its date, and the capacity statement is honest about one person’s time.
Illustrative example — not a client engagement. A consortium studying LLM-assisted code generation for scientific computing needs a correctness work package. The draft gives it three tasks: build an operator corpus with seeded faults and clean controls; run the partners’ generated kernels through seeded fuzzing against a high-precision reference; publish the corpus and a report. Dependencies: the use-case partner supplies the target operators by month three, the compute partner supplies GPU access. The capacity statement covers one researcher part-time, and the fallback says the corpus and method are still delivered if partner kernels arrive late.
Data rights and publication
Data rights, IP and publication terms come from the consortium agreement, which the partners’ institutions sign. I ask for them to cover publishing the method and non-confidential results; see publication, IP and independence.
Commission this if
- You lead or are forming a consortium, and one work package needs evaluation infrastructure, benchmark design, simulation or systems measurement.
- The consortium has its research question; the gap is someone who can build and run the technical part credibly.
- You, or your institution's research office, can check the call's eligibility rules for a non-institutional contributor.
Not the right fit if
- You need a partner to satisfy an eligibility requirement, such as a registered organisation of a given type or country. Nothing on this page establishes that.
- You need someone to sign the grant or consortium agreement, hold budget, or act as coordinator.
- The work package needs several full-time researchers for years and no other partner will staff it.
What you receive
- Work package description. Objectives, tasks and method, written so the lead partner can fit them into the proposal template.
- Dependency table. What the work package needs from other partners (data, access, use cases) and when; what it hands to other work packages and when.
- Deliverables and milestones. Concrete artefacts — a harness, a corpus, an evaluation report, a preprint — accepted on whether they exist and meet the agreed specification, not on what they show.
- Capacity statement. The effort one person can realistically commit over the project timeline, and what happens if the work grows beyond it.
- Risks and fallback. What the work package still delivers if a dependency arrives late or not at all.
What the study needs from you
- The call. The funding call text, its eligibility and participation rules, and the lead partner's reading of how a non-institutional contributor could take part, if at all.
- The consortium question. The research question and the technical gap you want filled.
- Partner roles. Who brings data, use cases, deployment environments and domain expertise.
- Dates. Submission deadline and expected project start.
How the work runs
- 1 Participation route first. The lead partner confirms whether and how an independent researcher can take part under this call. Calls differ, and I do not assume any route exists. If there is none, we stop here.
- 2 Scope the work package. Draft objectives, tasks and deliverables against the consortium question.
- 3 Dependency check. Walk the dependency table with the partners it names, so every input has an owner and a date.
- 4 Proposal text. I write the technical sections for my work package; the lead partner owns the proposal, the budget and every declaration.
- 5 Delivery, if funded. Deliver against the milestones; changes go through the consortium's own change process.
Limits and unfavourable results
- I am an independent researcher with no institutional signing authority. I cannot sign grant agreements, consortium agreements or declarations on behalf of any university or research organisation.
- Grant eligibility depends on the specific call and the lead partner. Nothing on this page, and no proposal text I draft, states or implies that I, or any organisation I am connected with, is eligible for funding.
- Capacity is that of one person. A work package that needs a team is only viable if other partners staff the rest.
- Drafting a work package does not commit me to delivering an unfunded project unless that is agreed separately.
Deliverables are artefacts and reports, not results in a chosen direction. If the work package shows that an approach does not hold under the tested conditions, that finding goes into the deliverable like any other.
Engagement terms
- Model
- A scoped technical work package inside a consortium proposal, through whatever participation route the call and the lead partner permit.
- Who does the work
- Dipankar Sarkar personally does the work package. Any specialist help is disclosed and agreed with the lead partner in advance.
- Commercial basis
- Set within the call's budget rules by the lead partner. Any fee is fixed in the work package and does not depend on the result.
- Availability
- Checked per call; the capacity statement is written for the specific project timeline.
Conflicts are checked before scoping. See publication, IP and independence.
Evidence behind this page
- The Correctness Illusion in LLM-Generated GPU Kernels
2026 · Preprint (not peer reviewed) · cited by 6 (Google Scholar, October 2026)
- Test-Input Generation for Tensor Programs: What Actually Finds Kernel Bugs
2026 · Preprint (not peer reviewed)
- Before the Pull Request: Mining Multi-Agent Coordination
2026 · Preprint (not peer reviewed)
- An Input-Regime Audit of Conflict Detection for Retrieval-Augmented Generation
2026 · Manuscript under submission
- gpuemu-corpus (GitHub) — Shared corpus with drivers for four papers, stored seeds and replay scripts: an example of a deliverable other work packages can reuse. MIT or Apache-2.0.
- ebrag-vecdb-2026-paper (GitHub) — Paper source, one JSON artefact per experiment and a runnable Python slice: an example of an audit deliverable. MIT.
Questions
What work package could a hands-on AI systems researcher credibly own?
One whose output is an instrument or a measurement: an evaluation harness or benchmark, an audit of a component across input regimes, or a coordination and logging substrate for multi-agent experiments. Each maps to work I have published as preprints or manuscripts, linked on this page.
Are you eligible for funding under our call?
I cannot answer that, and nothing here should be read as a yes. Eligibility depends on the specific call and the lead partner. I will provide the facts your research office needs about my status — independent researcher, no institutional signing authority — so they can decide.
Can you be listed as a partner organisation?
Only if the call and the lead partner confirm a route that allows it. I disclose my connections so they can be assessed: I founded Neul Labs, which builds AI agent infrastructure in Rust, and my research affiliation is Skelf Research. Neither is offered on this page as a consortium partner.
Who owns the results and publications?
The consortium agreement governs that. Before proposal text is final I ask for the right to publish the method and non-confidential results, with review windows for confidentiality rather than vetoes on findings.
See also
Collaboration
Industry-Academic AI Research Collaboration
Map a shared research question, split contributions and settle data and publication dependencies before the work starts.
Collaboration
Academic Seminars and Methods Exchanges
Research seminars and methods discussions for labs and seminar series, with an abstract, prerequisite reading and discussion questions.
Sponsored research
Publication, IP and Research Independence
A discussion checklist for IP, data handling, publication review, unfavourable results, authorship and conflict disclosure.
Commissioned study
AI Benchmark and Evaluation Study Design
Build a pre-specified test for a decision existing benchmarks do not measure, with baselines and failure conditions.
Send a research question
Include the call, its participation rules, the consortium question and the gap the work package would fill.
Last reviewed 2026-10-07.