Loading
Users set salary-day and balance rules in chat; Temporal workflows execute convert → transfer recipes with confirm cards and zero double-pays on webhook retries.
The Problem
A banking app offered crypto↔fiat conversion and payouts, but automation lived in brittle crons and support macros. Users wanted to say “when salary hits, convert and pay rent” — and have it actually fire once, every payday, without engineering tickets.
The Solution
We split Query vs Workflow agents, turned confirmed chat into Temporal RuleWorkflows, and bound salary / threshold triggers to Signals with idempotent Activities for convert, transfer, and notify steps.
Chat could answer balances and quotes (Query agent) or create durable automations (Workflow agent): salary credit → convert a share to crypto → pay fixed fiat payees → park remainder — with abort guards on minimum balance.
A single mega-agent mixed “what’s my balance?” with money movement tools and occasionally proposed transfers users never asked for. Splitting read-only Query from rule-authoring Workflow cut accidental side effects and kept confirm-card UX mandatory for any Temporal write.
Salary webhooks and balance threshold events start a short-lived RuleRunWorkflow (or Signal a supervisor that spawns one). Activities use ruleId + triggerEventId + stepIndex as idempotency keys so duplicate bank webhooks cannot double-convert or double-pay.
Each confirmed rule is a typed DSL document in the DB plus Temporal durable execution — not LLM memory. We use short RuleRunWorkflows per firing and Continue-As-New on any long-lived supervisor so we stay under Temporal’s ~51k event / ~50MB history hard limits. Blobs never go in Activity results; only ids.
Every new or edited rule rendered as a plain-language card (trigger, steps, caps, abort conditions) before commit. Edit and pause were first-class chat intents that mutated Temporal state — the LLM did not “remember” to run next month.
Beyond southbound tools into the ledger and conversion rail, we exposed the banking product itself as a northbound MCP server — get_balances, quote_convert, preview_salary_recipe, create_rule, pause_rule — so a future wealth or goals product could call the same Temporal-backed money path with scoped tokens, instead of re-integrating core banking.
Multi-step recipes handle partial failure (convert OK / transfer fail) with pause + support playbooks; in-flight pause is cooperative between steps; step-up auth on first/large money rules; payee screening and FX slippage caps live in Activities/DSL; webhooks are signed with replay windows; runs emit receipts (rule version + ledger ids) for support — not Temporal UI alone.
Business Impact
Payday recipes became a product surface, not a backlog item. Users author rules in natural language; Temporal owns waiting, retries, and replay when support asks “why did rent send?” Platform MCP means the next product plugs into the same money path instead of forking it.
Technical architecture write-upTech Stack
No. Workflow agent only persisted rules after the user confirmed the rule card. One-shot “run now” also required explicit confirm. Query agent stayed read-only or dry-run.
Each rule is a typed DSL record plus Temporal execution. We do not keep one immortal history forever: each payday/trigger is typically a short RuleRunWorkflow; supervisors Continue-As-New before Temporal’s history limits (~51k events / ~50MB).
The run marks PARTIAL, keeps a ledger-linked receipt, and follows a compensation / support playbook — it does not blindly re-run the convert step. The user can be notified; the rule can optionally auto-pause.
Ingress verifies signatures and a replay window, dedupes on provider eventId, then Activities key on the credit event id. The second delivery is a no-op.
Yes — that is the platform MCP design. Product B connects as an MCP client with scoped auth; create_rule and convert paths still hit the same Temporal Activities and policies. Partners never receive raw ledger credentials.
Same Temporal + agent discipline, different buyer: this is consumer/fintech product automation. The mid-market case study is maker-checker ops for KYC and payment repairs inside a bank.
Crypto / FinTech
Durable AI Ops Layer on Multi-Venue Exchanges
Desk copilots that turn natural-language intent into HITL-approved exchange orders — Vercel AI SDK, Temporal, and MCP — without YOLO trading bots.
Banking
Durable AI Ops Layer for Mid-Market Banking
KYC exceptions and payment repairs on Temporal + Vercel AI SDK + MCP — maker-checker Signals, PII redaction, zero double-posts on core retries.
Insurance
Claims Intake Routing & Data Extraction
FNOL packs auto-classified and routed — adjusters cut intake from ~45 minutes to ~5–8 on standard claims.