Skydo

International payments for Indian freelancers and SMBs fail silently: 38% of Skydo transactions experience delays or rejections (source: Q2'24 ops report), forcing users to manually decode SWIFT codes, chase support tickets, and reconcile banking messages. Each unresolved failure costs $87 in lost productivity and payment retries (source: 2023 RBI freelancer survey), while unresolved issues drive 22% churn within 90 days (source: churn cohort analysis).

Business case:
12,500 monthly failed payments × 38% auto-detectable issues × $87 savings = $413K/month recoverable value ($5M/year)
(Sources: Skydo transaction volume dashboard Aug'24, failure root-cause analysis, RBI survey)
If adoption is 40% of estimate: $2M/year
Downside: If false positive rate exceeds 5%, support costs increase by $120K/year (source: support cost model)

This IS an AI-powered diagnostic layer that classifies failures and prescribes resolutions in plain language. This IS NOT an automated payment reprocessing engine or compliance decision-maker. Given the $5M upside against $310K build cost (source: Regional Cost Benchmarks India), this warrants Q4 investment despite SWIFT message parsing risks.

Stripe automates retries but provides no root-cause analysis. PayPal offers generic failure codes without resolution steps. Wise explains failures but requires manual beneficiary updates.

WHO / JTBD: When an Indian freelancer's international payment fails, they want immediate clarity on why it failed and how to fix it – without deciphering banking jargon or waiting 48h for support.

BEFORE: Priya (freelance developer, Chennai) sees a $2,000 client payment marked "delayed." She spends 3.7 hours (source: UX research session avg) checking Skydo status, then her bank portal, then emails SWIFT MT103 codes to her cousin who works at HDFC Bank. They discover a beneficiary name mismatch ("Priya S." vs "Priya Sankaran"), requiring a corrected invoice and client reprocessing – delaying funds 8 days.

COST:

AFTER: Priya receives a Skydo notification: "Payment delayed: beneficiary name mismatch. Fix: 1) Update invoice to match bank ID 2) Client resends. Est. resolution: 2 days."

Data model:

class PaymentDiagnosis:  
  anomaly_score: float  # ML model output  
  failure_class: Enum  # BENEFICIARY_MISMATCH, RBI_COMPLIANCE_HOLD, FX_LIMIT  
  evidence: List[SWIFT_Field]  # e.g., ["Field 59: Priya S. vs Sankaran"]  
  resolution_steps: List[LocalizedStep]  
  eta_hours: int  

User flow:

1. Payment pipeline emits "delayed" event → Triggers anomaly model
2. Model parses SWIFT/RBI codes → Classifies failure
3. Generates diagnosis using RBI-compliant templates
4. Sends notification with "View details" CTA

Wireframe 1: Notification

┌───────────────────────────────────────────────┐
│ ⚠️ Payment Delayed: RBI Compliance Check       │
├───────────────────────────────────────────────┤
│ Client: TechGlobal Inc.                       │
│ Amount: $2,000                                │
│ Reason: Beneficiary name mismatch             │
│ Est. fix time: 48 hours                       │
├───────────────────────────────────────────────┤
│ [View Resolution Steps]          [Dismiss]    │
└───────────────────────────────────────────────┘

Wireframe 2: Diagnosis Dashboard

┌───────────────────────────────────────────────┐
│ Fix Payment: #TX-7843                         │
├───────────────────────────────────────────────┤
│ STEP 1: Update your invoice to match:         │
│   ▢ Full name: Priya SANKARAN                 │
│   ▢ Bank: HDFC Chennai Main (HDFC0000123)     │
│                                               │
│ STEP 2: Send updated invoice to client        │
│   [Email Template]        [Copy IBAN]         │
│                                               │
│ Impact: Client must resend payment            │
│ Estimated resolution: 48 hours                │
└───────────────────────────────────────────────┘

Phase 1 — MVP (6 weeks)
US#1: Auto-detect RBI compliance holds

• Given payment delayed >24h with SWIFT code "REJT / 6C"
• When system parses transaction
• Then classify as "RBI_COMPLIANCE_HOLD" with P0 accuracy (100% required)
• If fails: User misdiagnosis → legal risk
• Validated by: Compliance team against RBI handbook

US#2: Generate beneficiary mismatch resolution

• Given failure_class = BENEFICIARY_MISMATCH
• When user views details
• Then show exact field mismatch (e.g., "Field 59: Priya S vs Sankaran")
• And display "Update invoice" step with editable template
• Accuracy: P1 (≥99.5% field extraction)

Out of Scope (Phase 1):

Primary Metrics:

Guardrail Metrics:

Leading Indicator (D14):
If >40% of notifications get "View details" clicks → Predicts D90 resolution rate target (r=0.81 in pilot)

What We Are NOT Measuring:

• "Number of features used" (vanity; doesn't correlate with resolution speed)
• "Raw notification opens" (misleading; could indicate confusion)
• "Model confidence scores" (internal metric; irrelevant to outcomes)

Risk: RBI PA license amendment required for diagnosis advice
Probability: Medium | Impact: High
Mitigation: Legal team (A. Mehta) files Form D by 10/15; fallback: generic messages if not approved
Trigger: No RBI approval by 11/30 → Blocks launch

Risk: SWIFT message format changes break parser
Probability: Low | Impact: High
Mitigation: Monitor SWIFT alert feed (S. Patel); automated schema tests run hourly
Trigger: >5% parsing errors in prod

Risk: Users ignore resolution steps due to complexity
Probability: High | Impact: Medium
Mitigation: Embed "Copy invoice template" button; track step completion (PM: R. Verma)
Trigger: <30% step completion at D14

Risk: Correspondent banks block automated queries
Probability: Medium | Impact: High
Mitigation: Throttle checks to 5/min per partner; manual fallback protocol (Ops: P. Nair)
Trigger: >2 bank API suspensions

Kill Criteria (within 90 days):

1. False positive rate >7% (invalidates trust)
2. D90 resolution rate <45% (below baseline)
3. RBI issues compliance notice

4. 15% support ticket increase from confusion

Components:

1. Anomaly detector: Monitors payment status API → Flags delays/rejections
2. SWIFT parser: Extracts RBI-critical fields (59, 70, 72) using regex + NLP
3. Classifier: BERT model trained on 8K labeled failures (RBI codes = priority)
4. Resolution engine: Matches failure type to pre-approved RBI workflows

Assumptions vs Validated:

Decision: Depth of SWIFT message parsing
Choice Made: Parse only 8 critical RBI-mandated fields (59, 70, 72)
Rationale: Full MT103 parsing adds 3 months; RBI fields cover 92% of Indian cases (source: ICICI Bank data share). Rejected: Generic third-party parsers (inaccurate for RBI codes).

Decision: User notification timing
Choice Made: Notify only after 24h delay OR SWIFT error code
Rationale: Early alerts reduce float loss; 24h threshold avoids 73% of false positives (source: pilot). Rejected: Instant alerts on all delays (overwhelms users).

Decision: Resolution step automation
Choice Made: Manual user action (e.g., update invoice) → No auto-retry
Rationale: RBI prohibits automatic beneficiary changes. Rejected: Auto-retry for non-compliance cases (violates PSS Act).

Decision: Liability for incorrect diagnoses
Choice Made: Surface "This is not financial advice" disclaimer + support escalation path
Rationale: Limits regulatory risk; matches industry practice. Rejected: Guaranteed resolution ETA (creates contractual exposure).

Pre-Mortem:
"It is 6 months post-launch and this feature failed. The 3 most likely reasons are:

1. RBI blocked the feature over unlicensed 'advice' interpretation, forcing a 4-month rework of all messaging.
2. Users ignored resolution steps because the SWIFT field explanations were still too technical for non-bankers.
3. Wise shipped Hindi/English beneficiary self-service 8 weeks earlier, capturing 70% of our target cohort."

Success looks like: Priya's client pays in 48h after a failure. She tweets: "Skydo saved me 4 bank visits!" Support tickets for payment failures drop 65%. The CFO notes in the board memo: "This cut failure-related churn to single digits – the first time in 7 quarters."