SeedBridge

Inspiration

Founders outside the big startup hubs rarely know which grants or investors they're eligible for, so they write applications blind. Most rejections are avoidable: the wrong funder, the wrong stage, a missing number, a budget that doesn't add up. Funders face the mirror image, screening piles of unsuitable applications by hand. We wanted one place where both sides meet already prepared.

What it does

Founders describe their idea once in a five-step form, including where the business is now: just an idea, pre-seed, seed, Series A+ or a live business. SeedBridge then:

• scores funding readiness out of 100 across six dimensions, judged against what funders expect at that stage, with strengths, risks and the three fixes that matter most;
• lists what a panel will ask: the missing information and the questions to prepare for;
• builds a market snapshot from live web research, with sources;
• ranks matching funders by fit and flags eligibility problems before you waste time applying;
• writes a tailored application in one click, using only the founder's own facts. Anything missing is marked [ADD: …], never invented.

Funders list a grant, equity fund, prize, accelerator or loan, then get a ranked, pre-reviewed shortlist of founders who fit, each with the reason. They can shortlist, decline or request an intro.

How we built it

Seven AI steps, each on the smallest NVIDIA Nemotron model that does the job well, all served by Nebius Token Factory:

• RAG over a funding knowledge base we wrote (stage benchmarks, how grant panels score, what investors look for, writing guidance, raising in emerging markets). Each review lists the sections it was judged against.
• Matching is hybrid: hard eligibility rules (stage, sector, country, ticket size) and embedding similarity build a shortlist, then Nemotron Super reranks it and explains every match in one sentence.
• Behind the AI page: every model call is logged with model, latency, tokens and cost, so a full review is auditable and costs a fraction of a cent.
• Nebius Serverless: the app ships as a CPU container for a Serverless Endpoint, and re-matching runs as a Serverless Job.
• FastAPI, Jinja and SQLite with no front-end build step; Motion for animation. Model calls run as background tasks with live progress.

Challenges we ran into

• Reliable JSON from every model size. We parse defensively (fenced blocks, stray reasoning) and retry, so one malformed answer never breaks a review.
• Keeping drafts honest. Application drafts may only use the founder's facts; gaps become [ADD: …] placeholders.
• Right-sizing models. Running Ultra everywhere would be slow and expensive. Splitting the work across Lightning, Nano, Super and Ultra kept a full review fast and cheap without losing quality where it matters.
• Stage-aware judgement. A pre-seed idea and a trading business shouldn't be scored against the same bar, so the review retrieves benchmarks for the founder's stage first.

What we learned

• Small, fast models handle structuring and checklists well; the big model earns its cost only for judgement and long-form writing.
• Grounding with a short, well-written knowledge base improved consistency more than longer prompts did.
• Showing founders why (cited benchmarks, flagged eligibility rules) builds more trust than a bare score.

What's next

• Sinhala and Tamil reviews.
• Multi-slide PDF decks.
• A larger, verified catalogue of real funding programs.
• Real OAuth for LinkedIn, GitHub and Google, and funder analytics.

Built With

• css
• docker
• fastapi
• html
• httpx
• jinja
• motion
• nebius
• nebius-serverless
• nebius-token-factory
• nemotron-3-nano
• nemotron-3-super
• nemotron-3-ultra
• nemotron-3.5-lightning
• nemotron-3.5-super-vl
• nvidia-nemotron
• python
• qwen3-embedding
• rag
• sqlite
• tavily
• uvicorn