Fundvector

A multi-step funding agent that retrieves, verifies, and turns recommendations into actions.

Comment

Inspiration

For startup advisors and founders, manually sourcing non-dilutive funding is a massive time sink. It means spending countless hours sifting through fragmented government portals, parsing dense 50-page solicitation PDFs, and cross-referencing obscure eligibility criteria.

Worse, this manual grant discovery has too many hidden failure modes: an eligibility clause buried in long text, a deadline that changed, or a “global” call that still requires a local legal entity. Founders and advisors don’t need to waste hours building another generic link list — they need recommendations they can defend, generated instantly.

FundVector is built on one hard product rule: every surfaced claim must be source-linked or explicitly marked unverified. If we can’t prove it, we don’t present it as fact.

What it does

FundVector converts a startup profile into an evidence-first shortlist of non-dilutive funding options and a workflow plan you can execute immediately.

In the demo run shown in the app:

  • Input: US-based startup (AeroGrid Labs), seeking “non-dilutive funding for a climate resilience AI pilot”
  • Top match: NSF SBIR Phase I (Fit score: 92%, heuristic)
  • Deadline risk: Low urgency; 105 days remaining (Jun 05, 2026)
  • Immediate action: Review NSF SBIR solicitation topics for climate/AI alignment and begin preparing technical proposal by April 2026.
  • Additional ranked options: SUPERHOT SBIR/STTR and other alternatives.

FundVector also generates workflow tasks (owners + dates) and exposes execution actions (calendar export, email draft/send) so the output is operational, not just informational.

How we built it

FundVector Architecture Diagram FundVector is a multi-step agent application built explicitly with Elasticsearch + Elastic Agent Builder:

  • Search + index: Opportunities are indexed in Elasticsearch and retrieved with profile-driven search (structured fields + text).
  • Deterministic ranking: ES|QL + structured scoring keep the fit score stable and explainable under demo conditions.
  • Agent orchestration: Agent Builder runs a multi-step sequence: parse → search → reason → score → plan.
  • Trust contract: Claim-level provenance states (verified/unverified) with source links and evidence surfaced directly in the UI.
  • UX separation: The Results | Plan | Evidence layout keeps the main experience clean while preserving strict auditability.

The Run Pipeline shown in the product:

  1. Parse company profile
  2. Search funding index
  3. Run agent reasoning
  4. Score matches
  5. Assemble workflow plan
  6. Return response

Challenges we ran into

  • Trust UX was harder than ranking. We had to enforce a “no bluffing” rule without overwhelming users. Unproven claims must be prominently labeled unverified, and freshness/provenance must be visible in-context.
  • Reliability under live demo conditions. We added deterministic fallback paths and explicit capability gating so runs stay stable even if a step is slow or unavailable.

Accomplishments that we're proud of

  • Massive Time Savings: We turn hours of manual portal-scraping and PDF-reading into a fully automated run, allowing advisors to focus on application strategy rather than basic search.
  • Advisor-grade output: Ranked opportunities paired with immediate next actions, ready to hand off to a client.
  • Evidence-first UI: Inspectable claim provenance that solves the RAG hallucination problem.
  • Observable multi-step pipeline: Judges can follow the agent's logic end-to-end.
  • Execution readiness: In the featured demo run, FundVector reports 100% source-link coverage across surfaced recommendations and generates actionable workflow tasks.

What we learned

Trust only matters when enforced by product rules. Judges and users trust systems they can inspect: evidence, trace, and clear pipeline behavior will always beat black-box “AI magic.”

What's next for FundVector

  • Tighten verification semantics: e.g., “not specified” should never be shown as verified.
  • Expand structured extraction: Ensure more claims are provable rather than inferred.
  • Add evidence-gap workflows: Explain why something is unverified + one-click re-verify.
  • Progressive streaming: Stream results first, then progressively upgrade verification status live.

Built With

  • docker
  • docker-compose
  • elasticsearch
  • es|ql
  • fastapi
  • github-actions
  • google-artifact-registry
  • google-cloud-build
  • google-cloud-run
  • kibana-agent-builder
  • next.js
  • pydantic
  • python
  • react
  • recharts
  • smtp-(sendgrid/gmail)
  • tailwind-css
  • typescript
  • uvicorn
  • watcher
  • webhooks
Share this project:

Updates