Project summary

I am Manuel Castillejo Vela, a technical founder in Seville, Spain. With CTO Eduardo Robles Russo, I build FlowSight, a desktop workspace that uses local AI to help people recover work context and prepare useful summaries. We want to develop and release a first compact, task-specific open model/adaptation, with a local reference workflow and reproducible evaluation. Our own model has not yet been trained; the working app currently uses a third-party Qwen3-VL-2B model locally.

This grant would support public research and reusable outputs, available without purchasing FlowSight. It would not buy our private application code or subsidize marketing. The goal is to learn when small local models can provide useful assistance while keeping people in control, and to publish the evidence and limitations. We do not claim that local AI is always more efficient or already delivers carbon savings.

What are this project's goals? How will you achieve them?

Over ten weeks after funding, we would select a compact, license-compatible baseline and a bounded training or adaptation approach for editable work summaries. We would prepare artificial or explicitly permitted examples, record their provenance and separate training from evaluation. We would release the resulting model or adaptation, the relevant training/reference code, an evaluation set we have rights to publish, a model card and English/Spanish documentation under appropriate open licenses compatible with any upstream assets.

Weeks 1-2: define the task, reuse rights, baselines, evaluation protocol and compute budget. Weeks 3-6: implement and train/adapt a first candidate, logging the training configuration and resource use. Weeks 7-8: evaluate task quality, unsupported statements, instruction following, user-control boundaries, latency and memory on available desktop hardware. Weeks 9-10: obtain an independent review and publish a reproducible release, results and limits. Hardware electricity measurements would be reported only where we have reliable equipment; whole-system environmental impacts would not be inferred from partial measurements.

How will you know it worked?

Success means a downloadable first task-specific model/adaptation, a reference that runs locally, reproducible public tests, documented data/model provenance and a report that compares the candidate against the chosen baseline. We would report errors alongside performance. If adaptation does not improve useful task quality, we would publish that negative result and the tools rather than claim success. Training compute, unsupported hardware, privacy limits and evaluation weaknesses would be visible. We have not measured these outcomes yet.

How will this funding be used?

The maximum request is USD10,000: USD8,000 for 160 hours of model, reference and evaluation engineering at an indicative USD50/hour; USD1,000 for a capped training/compute budget; USD1,000 for independently contracted evaluation and release review, with the reviewer recruited after an award. No reviewer is already engaged. This is a scoped first release, not a budget to train a large foundation model from scratch.

At the USD5,000 minimum, we would reduce engineering to 80 hours/USD4,000, compute to USD500 and independent review to USD500. The smaller version would prioritize a license-compatible adaptation, local reference and fewer clearly documented test cases. No hardware purchases or paid growth are included. Actual supplier/compute choices would be made within the award and rights constraints; funds would be used for the agreed public outputs.

Who is on your team? What's your track record on similar projects?

Manuel Castillejo Vela is FlowSight Founder/CEO and a technical builder; Eduardo Robles Russo is CTO and a software engineer. Both work full-time in Seville and met as Software Engineering classmates. FlowSight operates through Manuel's Spanish sole-proprietor registration, dated 1 September 2026; a separate company has not yet been incorporated. Development began in 2025, with a public product launch in February 2026. We estimate 50-100 users in Spain and China; this is a founder estimate, not a measured monthly-active-user count. Revenue is zero. Existing local-inference experience supports the proposed engineering, but does not establish experience training a new foundation model.

Product: https://flowsight.site

Manuel: https://www.linkedin.com/in/manuel-castillejo-vela/

Eduardo: https://www.linkedin.com/in/eduardo-robles-russo-5133912b1

What are the most likely causes and outcomes if this project fails?

The baseline may be unsuitable, a compact model may not be accurate enough, adaptation may exceed the compute budget, or useful examples may lack clear release rights. We would narrow the task, cap spending and publish permitted code, tests and negative findings. We would not release data without rights or represent an upstream model as our original training. Broad adoption and reduced dependence on central inference are goals, not guaranteed outcomes. The open research outputs would remain usable independently of the commercial app.

How much money have you raised in the last 12 months, and from where?

We have received no external investment or cash grants. Programme memberships and recognitions are not cash funding. SECURE, FUTO, GitHub Secure Open Source Fund, Awesome Disability and the two initial Mozilla applications are pending. A separate USD1,000 Awesome Conservation and Climate application was just received for measurement equipment for a new public electricity experiment; no award is granted. There is also a proposed private fundraising round, not committed capital.

The Mozilla model/research proposals and this request are alternative routes wherever their scope overlaps. If more than one funder offers support, we would disclose it and resolve scopes before accepting awards; the same hours, compute and deliverables would not be charged twice. Any equipment funded by the climate proposal would not be billed here. A separate behavioral-safety pilot may also be submitted as alternative funding for overlapping evaluation work. All awards remain uncertain.

AI assistance was used to organize and translate my previously supplied project and personal information and this grant plan. I have not represented this as unaided human writing. This proposal describes planned work, not completed research.