From solicitation dropto submission-ready PDF.
Six agents. One workspace. Three live today, three in private beta. Below: what each one does, what it touches, what it doesn't, and what stays in your hands.
Six stages. One file, worked end-to-end.
Each row is a stage with a status tag. We won't pretend something is shipping when it isn't.
Discovery
LiveThe agent
Scans SAM.gov, Grants.gov, Simpler Grants, NY State, NYC DYCD, Candid, and SBIR.gov on a 6-hour cron. Fit score per opportunity: NAICS, keywords, geo, dollar band, prior funder history. ≥80 fit becomes a brief.
You
Open the morning brief. Three to five ranked opportunities. Click the fit, dismiss the rest.
Runs 24/7. ~3 min of your time per morning.
Vault
Private betaThe agent
Retrieves the artifacts that ground this proposal — past wins with the same funder, capacity statements, partner letters, logic models, 990s. Top-k chunks via pgvector, scoped to your tenant.
You
Confirm the retrieval. Add anything missing. Uploads stay in your tenant.
30 seconds to scan. Most users skip — it's right.
Draft
Private betaThe agent
Writes section by section in your voice fingerprint. Sonnet 4 by default; Opus 4.7 for evaluation-critical sections. Every claim is grounded in a vault artifact. Ungrounded claims get a [VERIFY] flag — never silently invented.
You
Read the draft. Resolve [VERIFY] flags by pointing to a vault doc or rewriting.
10–25 min per section vs. 2–4 hours from scratch.
Reviewer
Next this quarterThe agent
Opus 4.7 reads your draft against the funder's rubric. Scores 0–100. Emits specific revision requests with the rubric language cited. Brutal, not polite.
You
Accept what you agree with. Push back on what doesn't fit your strategy.
5–10 min per pass. Usually 1–2 passes.
Compliance
Next this quarterThe agent
Deterministic checks: page count, font, line spacing, margins, required attachments (W-9, audit, SF-424 family), budget arithmetic, indirect rate caps, eligibility, deadline timezone, 24-hour buffer.
You
Read the pass/fail report. Fix what's red. Re-run until clean.
Under a minute to run.
Submit
LiveThe agent
Assembles the final PDF/docx package, exports a clean submission bundle, writes the full agent activity log to your audit trail.
You
You submit. Always. The platform never auto-submits, never signs anything, never wires money.
Whenever you're ready.
Tuesday 9:14 AM. A $650K DYCD workforce solicitation drops.
Here's what happens by Friday — once all six agents are live. This is the target experience, not a shipped customer testimonial.
TUE 09:14
Discovery
DYCD posts a $650K Workforce Learning Grant for community-based orgs serving 18–24 in NYC. Scraper picks it up at the next 6-hour tick.
TUE 09:22
Match
Fit score: 94. NAICS aligns (624310). Geo aligns. Prior win with NYC HRA carries funder history weight.
TUE 09:28
Notify
Slack ping in #capture: 'New 94-fit. DYCD-WLG-2026. $650K. Due 5/29 17:00 ET.'
TUE 11:40
First drafts
Draft agent returns Need, Approach, Org Capacity, Logic Model in voice. Six [VERIFY] flags. Three resolved from vault, three flagged for staff to confirm.
WED 10:15
Reviewer pass
Opus 4.7 scores 78/100. Flags weak theory of change on Approach → Outcomes; missing evaluation timeline.
WED 14:30
Revisions
Writer accepts 7 of 9 revisions. Reviewer rescores 91/100.
THU 11:00
Compliance
Page count clean. Budget sums. Indirect at 10% — under the 12% cap. Three required attachments pulled from vault.
THU 16:20
Submit-ready
Final PDF assembled. Audit log exported. ED reviews, signs cover letter, submits via DYCD portal.
FRI 09:00
Filed
Confirmation logged against the proposal record. Eight days early. ~6 hours of staff time instead of 40+.
Total wall time → 4 days · Staff time → ~6 hours
Each one has a job. And a lane.
Six narrow agents beats one wide agent that quietly drifts. Each card names the job, what it touches, and where it is on the path to live.
Discovery agent
Always-on opportunity scanner across federal, state, city, and foundation sources.
Touches
SAM.gov, Grants.gov, Simpler Grants, NY State, NYC DYCD/HRA/DOE, Candid, SBIR.gov.
Never touches
Submit on your behalf. Pay subscriptions. Reach out to program officers.
Voice agent
Builds and refreshes your voice fingerprint from past wins, reports, founder letters.
Touches
Documents in your vault. JSON fingerprint on your org row.
Never touches
Train any model outside your tenant. Share your voice with other customers.
Vault agent
Retrieves the right artifacts from your tenant for grounding every section.
Touches
Past proposals, partner letters, logic models, 990s, capacity statements, bios.
Never touches
Read other tenants' vaults. Persist content outside your encrypted storage.
Drafting agent
Generates section-level drafts in your voice, grounded in your vault.
Touches
Section drafts, citations, [VERIFY] markers. Sonnet 4 default, Opus 4.7 for critical.
Never touches
Invent partners, fabricate outcomes, cite outside your vault.
Reviewer agent
Scores drafts against the funder's evaluation rubric and requests specific revisions.
Touches
Solicitation excerpts, extracted rubrics, your draft. Opus 4.7.
Never touches
Override your judgment. Edit your draft directly without approval.
Compliance agent
Deterministic pre-submit gate for page rules, budgets, attachments, eligibility, timing.
Touches
Final assembled package, solicitation requirements, SF-424 family.
Never touches
Submit. Sign anything. Make legal or eligibility judgments — it flags, you decide.
We get you to submit-ready. You decide what ships.
Some things should never be automated for federal capture. The line — and where we don't cross it.
Final submission
Human only
The last click is always yours. We assemble. You file.
Signing
Human only
No autopen, no signature stamp, no e-sign on your behalf.
Budget approval
Human only
We do the math. You approve the numbers.
Partner outreach
Human only
We surface partners. We don't email them.
Eligibility calls
Human only
We flag risks. We never say 'you qualify' to a federal officer.
Anything irreversible
Human only
If undoing it requires a phone call or a lawyer — you do it.
Your vault is yours.
Tenant isolation
Postgres RLS scopes every row to your org. Verified by automated tests on every deploy.
Voice fingerprint stays put
Voice JSON lives on your org row. Never used to train shared models, never pooled with other tenants.
No training pipelines
We don't fine-tune on customer data. Inference happens against your retrieved context at runtime.
Full security architecture, sub-processors, and compliance posture on the trust page.
Tenant ready in four minutes.
Not a sales call. Not a sandbox. Sign up, do these four things, run.
- 01
Create your org
Pick nonprofit, for-profit, or dual. Add NAICS codes, geo, funding categories you pursue.
60 seconds
- 02
Upload 5–10 vault docs
Past wins, annual reports, founder letters, capacity statements, partner letters.
90 seconds
- 03
Connect funding sources
Toggle SAM.gov, Grants.gov, DYCD, NY State, Candid, SBIR. We turn on the scrapers you need.
30 seconds
- 04
Run first Discovery
First scan kicks off. Real fit-scored brief in your inbox by tomorrow morning.
60 seconds + the wait