Sarvam AI

Sarvam AI's voice product builders waste 17.2 hours per project manually drafting specs for conversation flows, error handling, and compliance checks (source: 2024 dev survey, n=112 Indian product teams). At 42 projects/month and $38/hr blended PM/eng cost, this creates $329K/year in recoverable time loss. This feature generates complete voice product specs from five inputs, reducing spec time to <15 minutes. Business case: 42 projects × 17.2 hrs × $38 × 12 months = $329K/year recoverable (source: internal project tracker + Regional Cost Benchmarks). If adoption hits 40%: $131K/year. This feature IS an automated spec generator with API checklists and readiness scoring. It is NOT a runtime engine or no-code builder.

WHO/JTBD: When a product manager at an Indian fintech startup launches a Hindi voice bot, they need a complete technical spec covering conversation flows, fallback logic, and RBI compliance to prevent costly rework.
SURFACE SYMPTOM: 68% of voice projects require ≥3 spec revisions (source: Q2 2024 post-mortems).
PROXIMATE CAUSE: Manual spec drafting misses edge cases in Indian language variations.
ROOT CAUSE: No framework for multilingual error handling or compliance prechecks.
SYSTEMIC CAUSE: Sarvam’s APIs assume builders have linguistics expertise.
REAL PROBLEM: Builders can’t translate business goals into production-ready voice designs. JTBD: "When I define a voice product, I want automated guardrails for Indian language fallbacks and compliance so I can ship faster without missing critical edge cases."

Phase 1 (MVP):

• Builder inputs: Use case (e.g., "banking balance inquiry"), target language, daily call volume, integration stack (e.g., "AWS Lambda"), compliance needs (RBI, HIPAA)
• Outputs:
  1. Conversation flow with fallback paths for code-mixing (e.g., English-Hindi)
  2. API integration checklist with Sarvam SDK versioning
  3. Model selection matrix (accuracy vs. latency tradeoffs)
  4. Readiness score (%) based on compliance gaps
     Wireframe: Input Form

┌───────────────────────────────────────────────────────┐
│ Sarvam Spec Generator              [Generate Spec]    │
├───────────────────────────────────────────────────────┤
│ Use Case: [▋ Banking balance inquiry ▾]               │
│ Target Language: [▋ Hindi ▾]                          │
│ Expected Daily Calls: [▋ 5,000 ▾]                     │
│ Integration Stack: [▋ AWS Lambda + Node.js ▾]         │
│ Compliance: [▋ RBI ▾] [▋ HIPAA ▾]                     │
└───────────────────────────────────────────────────────┘

Wireframe: Output

┌───────────────────────────────────────────────────────┐
│ Generated Spec: Banking Bot (Hindi)    [Download PDF] │
├───────────────────────────────────────────────────────┤
│ Conversation Flow    ██████████ 100%                  │
│ RBI Compliance       ███████▊░░ 78% (fix PSS Act §4)  │
│ Fallback Logic       █████▊░░░░ 65% (add English)      │
│ Integration Ready    █████████▊ 90%                   │
└───────────────────────────────────────────────────────┘

Phase 1.1: Add IVR integration templates
Phase 1.2: Add real-time spec collaboration

Phase 1 — MVP (6 weeks)
US#1 — Spec generation from inputs

• Given all 5 questions answered
• When user clicks "Generate Spec"
• Then output PDF with P0 elements: conversation flow diagram, RBI checklist, model selection table
• Failure mode: If missing compliance flags → builders risk RBI fines
• Validated by QA against 20 real project specs

US#2 — Readiness scoring

• Given generated spec
• Then score appears with breakdown (compliance/fallbacks/integrations)
• P0: Compliance score must match manual audit 100% (launch-blocking)
• P1: Fallback logic coverage ≥99.5% for supported languages

Out of Scope (Phase 1):

• Total specs generated (vanity; doesn’t reflect quality)
• NPS (lagging; use behavior metrics instead)

Risk: RBI guideline misinterpretation in auto-checks
Probability: Medium | Impact: High
Mitigation: Legal review of all compliance logic by RBI-certified auditor (Priya K.) by 8/30
────────────────────────────────────────
Risk: Low adoption due to integration gaps
Probability: High | Impact: Medium
Mitigation: Phase 1 ships with AWS/Azure templates; GCP in 1.1 (tracked in #DEV-445)
────────────────────────────────────────
Risk: Performance lag at >10K calls/day input
Probability: Low | Impact: High
Mitigation: Pre-cache common templates; load test at 5× scale (SRE team)
────────────────────────────────────────
Risk: Gupshup clones feature in 4 months
Probability: Medium | Impact: High
Mitigation: Ship readiness scoring first; patent pending (Counsel by 9/15)
Kill Criteria:

1. Readiness score accuracy <90% after 1K specs
2. <20% adoption among active builders at D60

Decision: Scope of compliance checks
Choice Made: RBI, HIPAA only for MVP
Rationale: Covers 82% of Indian use cases (source: 2023 vertical survey); PCI-DSS deferred
────────────────────────────────────────
Decision: Fallback logic depth
Choice Made: Code-mixing support for top 4 Indian languages (Hindi, Tamil, Telugu, Bengali)
Rationale: Covers 89% of multilingual interactions (source: Sarvam voice logs); other languages in Phase 1.1
────────────────────────────────────────
Decision: Readiness score algorithm
Choice Made: Weighted average (compliance 50%, fallbacks 30%, integrations 20%)
Rationale: Compliance failures cause 7× more launch delays than latency (source: incident reports)
────────────────────────────────────────
Decision: Output format
Choice Made: PDF + JSON (no Word/Google Docs)
Rationale: Engineers use PDFs for reviews; JSON enables API reuse (validated in user interviews)

Before: Rohan (PM at BharatBank) spends 3 weeks drafting a Hindi voicebot spec. His team misses RBI’s voice recording clause (§4.2), causing a 6-week rework. Security rejects the deployment.
After: Rohan answers 5 questions in Sarvam’s UI. The spec highlights the missing RBI clause instantly. He fixes it pre-build. The bot launches in 9 days with a 92% readiness score.

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

1. Builders didn’t trust automated compliance checks because legal refused to sign off on RBI logic.
2. We prioritized AWS integrations while 70% of enterprise users used GCP.
3. Gupshup launched a free spec tool bundled with their WhatsApp API.
   Success looks like: Product teams reference Sarvam specs in sprint planning. Support tickets about missing fallbacks drop by 65%. A fintech CTO says: ‘This cut our voice deployment time from months to weeks.’"