queue / birdie-wellness / APP-0418 V3SANDBOX DW
Birdie Wellness, Inc.
Stripe incumbentMCC 5969DE C-Corp$9.99/wk subscription
87%
confidence · 412 comparables
Applications
14 today · Keel handled 11 — 3 need a human
MerchantScoreRecommendationStatusAge
Birdie Wellnesssendbirdie.com · MCC 5969
62K3
◆ Approve · conditions
Needs you · 3 itemsSLA 3h 48m
12m
Northwind Goodsnorthwindgoods.com · MCC 5969
74K4
◆ Decline drafted
Needs you · sign-offSLA 41m → M. Patel
41m
Cleo Fitnesscleofit.com · MCC 7298
55K3
◆ Approve · conditions
Needs you · 1 itemSLA 6h 02m
2h
Harbor Supplyharborsupply.co · MCC 5712
31K2
◆ Approved
✓ Auto-cleared
overnight
Brightpath Tutoringbrightpath.io · MCC 8299
14K1
◆ Approved
✓ Auto-cleared
overnight
Juniper Goodsjunipergoods.com · MCC 5947
28K2
◆ Approved
✓ Auto-cleared
overnight
Lumora Skinlumoraskin.com · MCC 5977
71K4
◆ Approved · reserve
◔ Monitoring · M3
3 wks
Showing 7 of 14 · auto-cleared files carry full ledgers — nothing is approved silently
Policy studio
Draft a policy change, then backtest it against your last 5,214 decisions before anyone goes live.
ACTIVE: 2026.Q2
Draft ruleschanges here touch nothing until approved
Auto-clear ceilingHighest tier Keel may approve without a human signature. Declines are never automated.
Minimum K3 reserveRolling reserve required for elevated-tier approvals.
5%
Refund policy gateRequire a published refund policy BEFORE activation (vs. within 30 days).
Max predicted dispute ratioFiles modeling above this go to manual review.
Backtest · last 12 months5,214 decisions replayed, outcomes knownNOT RUN
Approval rate
78.0%
baseline
Auto-clear rate
57.0%
baseline
Modeled annual loss
$312K
baseline
False declines
4.1%
baseline
Adjust the draft rules and run the backtest — you will see exactly what would have happened to last year's portfolio before a single live file is touched.
Scorecard weightsPATENT PENDINGour scoring logic — how much each risk domain moves the score — adjust, backtest, route for approvalΣ 100%
Weights are policy — they version, backtest, and require approval like any rule. Merchant-specific scorecards inherit from this base.
Calibration & learningevery decision and its 12-month outcome trains the model — your portfolio, not a generic one
5,214
decisions in your corpus — scored, signed, ledgered
3,981
with 12-mo outcomes attached (chargebacks, losses, churn)
+4.2pts
tier-accuracy gain vs. the base model, this quarter
91%
of "files like this" landed in tolerance — and it knows why
Every decisionscored against policy, signed, sealed to the ledger
Outcome attacheschargebacks, losses, churn flow back to the file at 30/60/90/180d
Model re-calibratesweights + thresholds refine; disagreements become labeled training signal
Next file is sharperbenchmarks, reserves and tiers tighten — backtested before they ship
Bring your back-book — past approvals, declines and their outcomes — and Keel calibrates to your portfolio from day one. No model trains on another customer's raw data.
Live: every signed file since you started is already feeding the next decision. Knowledge stops living in one underwriter's head.
Agents
Six governed agents work this queue. Each has a hard scope, every action is hash-chained, and none of them can decline a merchant — ever.
247 runs today100% ledgeredLearning: 14 overrides → challenger v2.7 in validation0 scope violations · 90d
Intake AgentNormalizes every application · chases missing fields
May: parse applications · message merchants · open filesNever: score · decide · contact banks
34 runs todaylast run 2m ago0 exceptions
Evidence AgentScreens, crawls, registries, reviews — every pull cited
May: run KYB/OFAC/crawls · attach cited evidence · flag conflictsNever: invent data (nulls + gaps only) · decide
96 runs todaylast run 40s ago0 exceptions
Chaser AgentNudges merchants on docs & connections · wakes snoozed files
May: send templated nudges · schedule follow-ups · auto-wake filesNever: promise outcomes · alter conditions
41 runs todaylast run 11m ago0 exceptions
Auto-Clear AgentApproves clean K1–K2 files inside policy, full ledger
May: approve K1–K2 within policy 2026.Q2 · attach controlsNever: decline · touch K3+ · approve with conflicts present
8 cleared overnightlast run 3h ago0 exceptions
Re-true AgentWatches conditions & events · reopens files on material change
May: monitor M-reviews, volume spikes, dispute drift · reopen filesNever: change tier or controls on its own
3 M6 reviews queuedlast run 1h ago0 exceptions
Decline DrafterDrafts decline memos + reason codes for human countersignature
May: draft memos · propose D-codes · cite the evidenceNever: send, record, or communicate a decline
1 draft awaiting youlast run 41m ago0 exceptions
Needs your sign-offagent work never ships itself2 PENDING
Decline Drafter · Northwind Goods memo
D-302 Misrepresentation + D-304 laundering indicators, 6 evidence citations attached. Needs your countersignature to take effect.
Auto-Clear Agent · scope change request
Requests ceiling raise to include K3 files scoring ≤45 with zero conflicts. Backtest attached: +11% auto-clear, +$6K modeled loss. Requires Risk Director approval.
✓ Nothing waiting on you
Agents keep working — anything needing judgment lands here.
Agent activitylive · every row hash-chained
AgentActionFileWhenHash
EvidenceCrawl + 7 screens complete, 14 citationsCleo Fitness40s9e2f…
IntakeApplication normalized · 2 fields requestedJuniper Goods2m71ac…
ChaserRefund-policy nudge #2 sent · follow-up +3dBirdie Wellness11mb8d1…
Re-trueM3 reserve review opened on dispute drift 0.56%Lumora Skin1h44f7…
Auto-ClearApproved K2 within policy · controls attachedHarbor Supply3hc05e…
Decline DrafterMemo drafted · D-302/D-304 · 6 citationsNorthwind Goods41me3b9…
Pausing an agent, changing its scope, or overriding its output is itself a ledgered event — the agents are governed by the same chain they write to.
Integrations
Bring your own risk stack. Keel doesn’t resell data — your KYB/KYC vendors feed evidence, and Keel turns it into a defensible decision with the source cited on every row.
7 connected412 pulls todayavg 1.8s
Your stackconnect once — every file uses it automatically
Coverage by risk domainwhere your stack feeds the decision — and where Keel’s own research fills in
Gaps aren’t blockers — Keel’s Evidence Agent covers them with cited web research — but a connected vendor upgrades those rows from AI-inferred to verified-by-source.
Processing history · parsed from Stripe + statementsOCR + spreading, every figure cross-checked1 conflict
MonthVolumeTxnsAvg ticketCB%Refund%Flags
Feb$24.1K2,410$10.000.29%2.1%clean
Mar$28.7K2,861$10.030.35%2.4%clean
Apr$31.2K3,098$10.070.41%3.0%refund ↑
May$38.9K3,852$10.100.43%3.2%NSF ×1
Ticket mix: 96% low · 4% mid · 0% high ✓ Descriptor stable ✓ MCC drift: none ✓ Refund trend +1.1pt / 90d Conflict: stated $50K/mo vs observed $38.9K — flagged to evidence
Ownership · UBO treeCIP captured · all owners verified
Birdie Wellness, Inc.DE C-Corp · EIN •••4821
Mara Patel-SinghCEO · 62%CIP ✓ · ID ✓
Daniel OkaforCOO · 28%CIP ✓ · ID ✓
Employee poolESOP · 10%<25% · exempt
No layered entities, trusts, or foreign control. Control person matches signer. Changes to ownership trigger re-underwriting automatically.
Bank statements · OCR + spreading3 months parsed · tamper check passed
MonthDepositsEnd balanceNSFCross-check vs processingTamper scan
Mar$29.4K$18.2K0matches ±3%clean
Apr$32.0K$21.7K0matches ±3%clean
May$37.1K$19.4K1matches ±5%clean
Deposits reconcile with processing volume — the stated $50K/mo on the application does not. That conflict is written to Evidence as a Conflicting-source row and feeds judgment call #1 territory: clerical optimism vs. misrepresentation.
Cleared today
12 +8 auto
▲ best this month
Median decision
6m 12s
▼ 18% vs last wk
Approval rate
78%
— flat · policy 2026.Q2
Auto-clear rate
57%
▲ 9 pts this quarter
Portfolio VAMP
0.41%
limit 0.65% · 4 qtrs in tolerance
Decisions · last 14 days human-signed   auto-cleared
MAY 28JUN 3JUN 10
Tier mix · live portfolio1,284 merchants
1.3k ACTIVE
K1–K2 · 44%
K3 · 30%
K4 · 17%
K5 · 8% · enhanced monitoring
Time to decision · trailing 8 weeks▼ 41% since parallel-run
10m 24s 6m 12s
Queue agingSLA: decision-ready < 24h
< 1h
9
1–4h
4
4–24h
1
> 24h
0 ✓
Portfolio watch · top exposuremonitored against VAMP / ECP continuouslyall in tolerance
MerchantTierMo. volumeDispute ratioStatus
Lumora Skin · MCC 5977K4$1.9M0.56%Watch · M3 review
Birdie Wellness · MCC 5969K3$350K0.43%Pending commit
Cleo Fitness · MCC 7298K3$610K0.31%Healthy
Harbor Supply · MCC 5712K2$2.4M0.18%Healthy
Northwind Goods · MCC 5969K4 · pendingn/aDecline drafted
Team throughput · today
DW
David Whitely12 cleared · 6m median
★ best
MP
M. Patel9 cleared · 8m median
JL
J. Liu7 cleared · 7m median
KEEL
Keel · auto-clear8 files · full ledgers · 0 exceptions
57%
0
K3 · Elevatedexposure, not findings
re-true at month 6
Recommendation
◆ Approve
with 4 conditions
Reserve
5% · 90d
Loss coverage
8.8×
Keel did the work
94% 541 of 576 checks automated · 3 judgment calls are yours
Your judgment callseverything else is handled1–3 to act
1
Tax ID name mismatch — clerical?
IRS letter and application differ by suffix only. Model reads clerical, not material.
CP-575 “Birdie Wellness, Inc.” → Application “Birdie Wellness Inc.
✓ Confirmed clerical — written to ledger
2
Refund policy unpublishedOVERDUE 5D
Approval condition. A firmer nudge with a deadline is drafted and ready to send.
✓ Nudge sent — follow-up scheduled +3d
3
M. Patel countersignature
Required for K3. She signs 9 of 10 within 4 hours of a ping.
✓ Pinged — typically signs within 4h
✓ Queue clear — ready to sign
All judgment calls handled. Press S to sign + commit.
Why this decision?+
In plain English: this is a healthy young company in a risky billing category. Screens are clean ×7 and ownership is verified — the score is driven by where they sell (MCC 5969) and how they bill (weekly subscription), not by anything bad they did. A 5% reserve means even a bad dispute year costs the portfolio nothing (8.8× covered). That is why the model says approve with conditions instead of clean approve or decline.
What would change this decision?+
To clean Approve (K2): published refund policy (−6 pts) + 3 months disputes <0.5% (−5 pts) → score ~51. Automatic at the month-6 re-true.
To Decline: confirmed transaction laundering signals, sanctions hit, or dispute ratio >1.0% in first 90 days would re-tier K5 / prohibited.
To bigger reserve: drag the what-if slider — at 7% coverage hits 12.3×, which clears the sponsor-bank overlay threshold.
Site & storefront verification660 Market St, San Francisco — matches formation docs & bank referenceAddress verified
Aerial NW Aerial NE Aerial SW Aerial SE
AERIAL VIEW
OFFLINE — TILES UNAVAILABLE
Aerial view
SATELLITE · Z17
✓ commercial bldg
MAP VIEW
OFFLINE — TILES UNAVAILABLE
Map view
FINANCIAL DISTRICT · SF
✓ matches docs
Exterior close-up
EXTERIOR CLOSE-UP
OFFLINE — TILES UNAVAILABLE
Exterior close-up
EARTH IMAGERY · Z19
✓ active site
Feel better,
every week.
Website snapshot
SENDBIRDIE.COM · MAY 6 14:23
! no refund policy
Files like this onen = 412
91% performed in tolerance
Cohort: subscription · MCC 5969 · seed-stage. This file models at the 84th percentile of dispute risk.
Cohort median CB
0.38%
12-mo in tolerance
91%
Typical tier
K3 · 58%
Reserve norm
4.6%
Reserve what-ifLIVE
0%5% · 90 days10%
Coverage
8.8×
Merchant drag
Low
Exp. loss
$23.9K
Evidence · cited & snapshottedevery row hashed at pull time
Enriched by your stack:
FindingSeveritySource
OFAC, MATCH, FinCEN BO, card-brand registers — clear ×7ClearREGISTERS · MAY 6 14:22
Subscription billing on MCC 5969 — disputes 1.4–2.1× baseline historicallyMaterialPORTFOLIO 2023–26
No enforceable refund policy; /faq references “30-day satisfaction”NotableCRAWL · /FAQ /TERMS
4 billing-surprise complaints in 90 daysNotableTRUSTPILOT · IDS ON FILE
Tax ID name suffix mismatch (CP-575 vs application)ReviewIRS CP-575 · 04-25
Founders LinkedIn-verified; $2.4M seed (First Round + 4 angels)ClearSOS · LINKEDIN · MAY 6
BBB A− · accredited since 2023 · no unresolved complaintsClearBBB · MAY 6
Prior founder venture closed Q3 2022, reasons undisclosedNotableDE SOS FILING
Explainability576 signals · 12 domains · every factor attributed & cited
FactorWeightDirection
MCC / industry risk.92↑ risk
Subscription billing.84↑ risk
Chargeback exposure.76↑ risk
Refund policy gap.52↑ risk
Founder history.40↑ risk
Processor relationship.30↓ risk
Ownership transparency.22↓ risk
{ "score": 62, "tier": "K3", "recommendation": "approve_with_conditions", "reserve_pct": 0.05, "reserve_period_days": 90, "model": "keel-uw-2", "policy": "2026.Q2", // versioned, diffable "prompt_hash": "sha256:e3f0…1a4b", "mode": "recommend" // never "auto" }
Decision ledgertamper-evidentSHA-256 hash-chained at write time
T+ActorEventHash
00:00systemApplication opened · intake structured7c1d…
00:48screenersKYB / OFAC / MATCH / PEP / sanctions / media — clear ×7a9e2…
01:12crawlerSite crawl — 4 pages snapshotted3f8b…
02:51modelScored 62 / K3 · 576 signals · policy 2026.Q2d04a…
14:31D. WhitelyAnalyst signed memo of recordb6c7…
LEDGER ROOT SHA256:F2A8…91CD
Signature chain · K3 requires 4enforced in code — no auto-commit path exists
RoleWhoVoteStatus
AnalystDavid WhitelyApprove · conditionsSigned · 14:31
Senior UnderwriterM. PatelPending
Risk DirectorAuto-routed after Sr. UWQueued
Sponsor BankFirst Federal AcceptanceNotify on commit
Votes are recorded per signer — a split committee (any Decline vote on K4+) auto-escalates to the Risk Director with both rationales attached.
You are signing as a credentialed human decision-maker. The model's output is advisory; your signature is hashed to the ledger and replayable in audit, indefinitely.

Sign + commit decision

Approve with conditions · K3 · 5% reserve / 90-day · APP-0418
    Decision committed
    Signature 3 of 4 hashed to ledger · routed to Risk Director
    Re-score cadence attached: 30/60/90/180d · Sponsor bank notified

    New decision file

    Keel researches the merchant live — website, BBB, reviews, registries — and drafts the file.

    Send the application

    The merchant gets a magic link to the online application — business info, account connections (Stripe, Plaid, Shopify…), documents. You see progress live in the queue.
    Works standalone too — share it any way you like. Expires in 14 days; progress autosaves for the merchant.

    API key

    Live underwriting calls the Anthropic API directly from your browser, using Claude's built-in web search.
    Stored only in this browser (localStorage) and sent only to api.anthropic.com. Get a key at console.anthropic.com → API keys. Web search must be enabled for your org. Each run ≈ a few cents.

    Request data connections

    Pick what to ask the merchant for — each one strengthens evidence and speeds the decision.
    KEEL ACADEMY
    Learn the cockpit. Earn the streak.
    Self-guided tours that drive the real app — nothing to read, everything to click. Pick a track:
    ▶ Watch: Keel in 90 secondsSELF-PLAYING · NO SOUND NEEDED · 7 SCENES
    FILM
    OR TAKE A HANDS-ON TRACK — DRIVES THE REAL APP
    Together the tracks cover every surface: queue, decision file, judgment calls, evidence & provenance, financials, site verification, cohort benchmarks, the shared file, live AI runs, Policy Studio, scorecard, backtests, agents, integrations, merchant connections, invite flow, snooze/auto-wake, signatures & committee, ledger, dark mode & mobile.
    ESC TO EXIT ANYTIME · ←→ TO NAVIGATE · COMPLETIONS ARE LOGGED TO YOUR LEDGER (YES, REALLY)
    1 / 7