ARCHITECTURE

Designed like infrastructure, not a demo.

design — in development

LangGraph is the brain, Temporal is the nervous system. Every step of a decision checkpoints into a durable, replayable history — so an approval can wait a month at zero compute cost, and a failure mid-write never leaves a half-posted order.

THE TWO LANES

A brain and a nervous system.

LangGraph — the brain

The decision logic is an explicit graph: extract, check, simulate, recommend, gate, write. Each node is small, testable, and inspectable — no monolithic prompt doing everything.

Temporal — the nervous system

Every execution is durable: checkpoints at every node, event history that replays byte-for-byte, and saga compensation that walks backward cleanly when a step fails.

Temporal is the platform behind OpenAI's Codex work and JPMorgan's payments modernization — in their own published talks. Corin's orchestration is designed on it; we say 'designed', not 'built', until it's running.

FAILURE, BY DESIGN

The compensation path is a first-class citizen.

Write quote → reserve stock → post order → notify. If step three fails, steps two and one unwind in order and the ledger shows the whole walk. Most automation demos never show you this arrow. It's the most important one.

WRITE QUOTERESERVE STOCKPOST ORDERNOTIFYSTEP 3 FAILSCOMPENSATE — UNDO 2 · UNDO 1
SAGA — WRITE PATH + COMPENSATIONFIG 01
REPLAYABLE HISTORY

Any decision, replayed exactly.

Because every execution is event-sourced, any past decision can be re-run deterministically — same inputs, same versions, same output. For auditors, for debugging, for trust.

timeactoractiongatestatus
14:02:11m.chenQuote Q-2291 approved — execution 01J9… beganhumanapproved
14:02:11temporalCheckpoint: quote written, stock reservedevent history replayableautoapproved
14:02:13temporalPost order failed — compensation walked 2 · 1no half-written orderpolicypending
DATA BOUNDARY

Your ERP stays the system of record.

Reads are scoped

Corin reads through released, permissioned APIs — the same contracts your integration team would sign.

Writes are confirmed

Nothing writes below the autonomy level you set, and every write carries the approval that authorized it.

No training on your data

Your operational data is never used to train models. Learning happens inside your tenant, on your outcomes.

DECISIONS, WITH RECEIPTS

See a decision simulated before it's made.