Custom engagements & applied research

Custom AI engagements

The standard research agent solves one shape of problem well — a question, the public literature, a cited Markdown report back. Plenty of useful work doesn't fit that shape: your data lives behind a license the agent can't touch, your domain needs careful tuning, your pipeline expects a different output, or the workflow is more than a single query at a time.

Engagements we have run or are scoping conversations about:

• Connecting the agent to proprietary databases, internal corpora, or paywalled sources your group already pays for.

• Tuning agent prompts, retrieval, and tool sets for a specific scientific domain or therapeutic area.

• Building bespoke evaluation harnesses so a team can measure agent quality on the questions that actually matter to their pipeline.

• Designing institutional workflows — grant drafts, tumor boards, regulatory submissions, lab-notebook integration — where the agent is one component of a larger system.

We don't pretend to do everything. The strongest engagements are ones where the underlying problem is a research-or-literature problem we have built deep machinery for. If that sounds like the problem you're working on, write us at info@sparkit.science with a paragraph about it. We respond with whether it's a fit, what a scope might look like, and a rough timeline.

Applied research

SPARKIT is built by a team that also runs applied research on agents, evaluation methodology, and AI for science. The HLE-Gold and GAIA numbers on the home page came out of that work; future evaluations and methodology notes land on the public blog rather than stay internal.

Current and recent threads:

• Benchmarking the agent on Humanity's Last Exam (gold subset) and GAIA against frontier models and search APIs.

• Evaluation design for scientific reasoning where ground truth isn't a single number.

• Agent safety work — refusal training, output screening, citation-fidelity audits.

• Public write-ups of real-world workflows (CRISPR-screen triage, hereditary-cancer panels) on the blog.

If you're working on related research and want to collaborate, or want to read the methodology behind a specific evaluation, reach out.

What's coming

We have a small backlog of services we're scoping for specific corners of the research-tooling space. If your use case doesn't fit cleanly into the off-the-shelf agent or a custom engagement as described above, we still want to hear about it — it may shape what we build next. Write us with a paragraph about the problem.