In plain terms, this stage searches current funding sources, validates source links, compares eligibility against your org profile, flags risk or questionable output, organizes approved leads, and supports evidence-grounded proposal drafting. The table below is the technical detail behind that: which model runs each step, and its usage limit.
Model
Role & function
Limit
groq/compound
Layer 1 · DiscoverLive web search, up to 10 searches/call. Returns funders, amounts, deadlines, URLs. Memory excludes known funders.
groq/compound
Layer 1.5 · Currently bypassedLive per-URL checking is temporarily disabled, it was dropping valid URLs on the free tier. Every result is marked "needs check," confirm the link yourself before applying.
gpt-oss-120b
Layer 2 · QualifyScores 1–5 vs org profile. Mission, geography, eligibility, risks. Returns structured JSON with recommended action.
gpt-oss-safeguard
Layer 2.5 · GuardLanguage, bias, accuracy, appropriateness. Soft-flag only, AI never hard-blocks. Humans decide.
Google NotebookLM
Layer 4 · ProposeDrafts from real org documents only. Prompts instruct the model to use only uploaded evidence. Marks gaps [INSERT: ...]. Upload once, reuse forever.
One Groq key covers all AI layers. Weekly run uses ~16,200 tokens, under 10% of the free daily limit. Resets every 24 hours.