ai ops layer for nonprofits. one chat unifies 11+ saas tools (salesforce, bloomerang, quickbooks, zoom) into cited, plain-english answers. live with rivertown community foundation ($2.4m budget).
one chat, eleven saas tools
nonprofit teams run their work across 11+ disconnected systems — bloomerang, salesforce npsp, microsoft 365, sharepoint, quickbooks, instrumentl, zoom, power automate, power bi, knowbe4, plus solana for treasury. donors live in three of them. grants in two. finance in another. nobody knows where the truth lives, so the answer is always "let me get back to you."
kali is the agentic context layer that sits on top. one chat. plain-english questions across the whole stack. every answer cited back to its source record. no migration, no rip-and-replace.
the bar: four minutes to first answer
"in four minutes i knew which 14 lapsed donors to call this week, and why each one. used to take my dev director half a day." — sarah chen, executive director, rivertown community foundation
new orgs onboard in six steps: account → org profile → pick stack → oauth each tool (~10s each, read-only) → drop historical data → land on the dashboard. by minute four kali is answering questions about your nonprofit using your real records.
what's wired
eleven first-class connectors — each one exposes typed query functions to the agent plus a real oauth integration. cross-tool entity resolution unifies the same donor across bloomerang + salesforce + m365 + zoom into one canonical kali id, so the agent can chain getDonor → getRelatedAccount → searchEmail without us writing a join.
citation chains — every claim kali makes is marked [N] inline. the frontend resolves each [N] to a clickable chip → opens the source record. no black boxes.
onchain rails — x402 agent donations: every tenant gets a public http 402 endpoint, any ai agent can pay usdc over the wire to a nonprofit, receipts auto-issued. cause coins: per-tenant spl token-2022 mint with onchain metadata, 1b supply to the treasury.
persistent chat + audit log — every conversation, every tool call, every onchain settlement recorded immutably. row-level security in postgres. open any thread from a month ago and the source-pulse + receipts replay exactly as they happened.
live with rivertown
deployed in production with rivertown community foundation — sacramento-based, $2.4m budget, six active programs:
- 863 people resolved across systems
- 2,437 donations indexed
- $5.2m lifetime giving tracked
- 38 active grants in pipeline
- 60 zoom meetings + 30 transcripts indexed
- $303k usdc disbursed on solana (devnet, mainnet q3)
- 70+ tools across 11 connectors
- avg query latency: 2.6s
the stack
next.js 16 app router (turbopack) + typescript strict + tailwind 4 on the frontend. claude sonnet 4.6 with parallel tool-use + prompt caching (~90% input-token savings) across 70+ tools. supabase (postgres + auth + rls) for persistence. openai text-embedding-3-small over ~7k record chunks per tenant, hybrid retrieval (semantic + structured filters), top-k=20 with reranking. solana web3.js + spl token-2022 with x402 facilitators (payai / coinbase cdp), privy delegated signing for tenant treasuries. soc 2 type 1 in progress, irs 990 + aml/sanctions screening on every settlement.
what shipped
hackdavis 2026 sweep — best use of solana + best ui/ux + best hack for social good + most challenging hack (4 categories). $40k raised to date ($30k angel + $10k openai grant). featured in the chronicle of philanthropy as "ai tools nonprofits actually want." backed by the open society foundation family stabilization fund.
team: stephen hung (lead) + matthew kim + silas wu + jake li. trajectory: yc.
