TOGAF ADM — The Autonomous HR

TOGAF ADM — Phase A

Architecture Vision.

Phase A defines the scope, constraints, and measurable target state before any design work begins. It answers one question: what does success look like, and how will we know when we've achieved it?

Architecture Vision Statement

"By Q4 2026, 90% of all HR interactions at Rathi Textiles — leave applications, balance queries, policy questions, and payslip delivery — will be handled autonomously without human intervention, in under 60 seconds per interaction, across all six languages spoken by the workforce, at a total infrastructure cost below ₹2,000 per month."

The channel meets the worker, not the other way around

No new app. No new login. No training required. Every employee interaction happens on WhatsApp or a voice call — channels already embedded in daily life. Any solution that requires onboarding has already failed.

↳ Architecture constraint: Zero new client-side software

Policy governs. Code executes. Humans override.

All HR decision logic derives from the HR Policy PDF. No business rules are hardcoded. When the policy changes, the behaviour changes — without a developer, without a deployment, without a ticket.

↳ Architecture constraint: RAG-governed decision engine

III

Human oversight is designed in, not bolted on

HITL is not a fallback. It is a first-class architectural state with defined entry conditions, a presentation contract, a timeout behaviour, and an immutable audit record. Every escalation is a designed mechanism.

↳ Architecture constraint: HITL on confidence < 0.8

Immutable audit trail before any state mutation

Every agent action writes to an append-only audit log before any downstream effect. The audit trail is a by-product of normal operations — not a separate compliance process. It is the defence in a labour dispute.

↳ Architecture constraint: Firestore append-only rules

Serverless by default — zero idle cost

A 52-person textile business does not generate HR interactions at 3am. Every service scales to zero when idle. Cost aligns with actual usage. An always-on server is an architectural failure for this customer segment.

↳ Architecture constraint: Cloud Run scale-to-zero

OSS at the inference edge — GCP at the control plane

The customer already has GCP. The architecture keeps GCP for what it does best: managed identity, eventing, and LLM inference. OSS (Whisper, IndicTrans2, pgvector) replaces managed AI services only where accuracy and latency guarantees hold.

↳ Architecture constraint: Strategic hybrid — not wholesale replacement

Stakeholder concern → Architecture response mapping

Stakeholder	Primary concern	Architecture response
Priya Rathi	Cost below ₹2,000/month, Sunday payroll reclaimed	Serverless scale-to-zero. Payroll automation in Phase 3. Total infra: ~₹550/month.
Floor workers	Confirmation in their language in < 60 seconds	Whisper STT + NLLB-200 translation. Multi-language response generation via Gemini Flash.
CA / Accountant	Accurate payroll data, ESIC/EPF compliance	Immutable Firestore ledger. Monthly export in Tally-compatible format. Statutory register auto-maintained.
Labour Inspector	Leave register per Shops Act, evidence of fair treatment	Append-only audit log with policy clause citation on every decision. 7-year retention.

TOGAF ADM — Phase B

Business Architecture.

Phase B defines the target business capability model — what the organisation must be able to do, and how the architecture closes the gap between current state and target state. Every capability maps directly to a pain point identified in the client brief.

Capability 01 · Phase 1
Leave Application & Approval
AS-IS: WhatsApp to supervisor, unrecorded
Worker sends voice note or message. System transcribes, checks policy and balance, approves or denies in < 60 seconds. HITL fires on ambiguity. Every decision is recorded with policy clause and confidence score.
Full automation — 90%

Capability 02 · Phase 1

Leave Balance Query

AS-IS: Ask supervisor verbally, often wrong

Worker asks via WhatsApp text or voice. System queries Firestore and responds in real time with casual, sick, and earned leave balances in the worker's language.

Full automation — 100%

Capability 03 · Phase 1

Policy Q&A

AS-IS: Ask supervisor, answer unreliable

Worker asks any policy question in any language. Policy RAG retrieves the governing clause and responds accurately, citing the source paragraph. Every answer is consistent and auditable.

Full automation — 95%

Capability 04 · Phase 2

Employee Onboarding

AS-IS: Paper forms, incomplete, ESIC delayed

Onboarding agent collects Aadhaar, PAN, bank details, emergency contact via WhatsApp conversation. Triggers ESIC/EPF registration. Delivers digital offer letter. Compliance from day one.

Assisted automation — 80%

Capability 05 · Phase 3

Attendance & Payroll

AS-IS: 4-hour Sunday manual process

WhatsApp check-in/out. Automated attendance ledger. Payroll calculation engine. Payslip delivered as WhatsApp PDF. Payment disbursement trigger to Priya's accounting workflow.

Full automation — 85%

Capability 06 · Phase 4

Grievance Management

AS-IS: Verbal, unrecorded, unresolved

Worker logs grievance via voice or text. Statutory register maintained. Escalation workflow routes to Priya. Resolution tracked. Labour inspection export generated on demand.

Assisted automation — 70%

Business process — leave management (target state)

Step 1

Worker initiates

Voice note or text on WhatsApp

↳ Worker speaks naturally in any language · No script or menu · Voice or text

Step 2

System transcribes

Whisper STT + language normalisation

↳ Audio → transcript → structured intent · Language detected · < 4 seconds

Step 3

Policy checked

RAG lookup + balance verification

↳ HR Policy PDF queried · Leave balance retrieved · Confidence scored

Step 4

Decision made

Auto-approve / deny — or HITL

↳ Confidence ≥ 0.8: autonomous decision · < 0.8: owner notified via WhatsApp

Step 5

Audit + confirm

Immutable log → WhatsApp confirmation

↳ Audit log written · Worker notified in detected language · Total: < 60 seconds

Key performance indicators — Phase 1 target

End-to-end leave decision SLA < 60s

Autonomous decision rate (no HITL) ≥ 90%

Languages supported (auto-detected) 200+

Monthly infrastructure cost ≤ ₹550

New apps required for employees 0

STT confidence threshold (auto-HITL) 0.85

Policy RAG confidence threshold (HITL) 0.80

TOGAF ADM — Phase C

Information Systems Architecture.

Phase C defines the application architecture and data architecture — what systems exist, what they do, and what data they manage. This is split into two sub-phases: Application Architecture (what runs) and Data Architecture (what is stored).

Phase C1 · Application Architecture

Channel Layer

WhatsApp Business API · IVR Voice Gateway

↳ Meta WhatsApp Cloud API — receives inbound audio/text, sends outbound confirmations and HITL alerts
↳ Exotel / Plivo SIP — handles inbound IVR calls, streams audio to STT service via WebSocket
↳ Both channels are abstracted behind a gateway interface — channel layer is provider-agnostic

Speech Layer

Whisper large-v3 · NLLB-200 (Meta OSS)

↳ Whisper: audio → transcript, language detection, confidence score — containerised on Cloud Run GPU spot
↳ NLLB-200: normalises code-switched speech (Hinglish, Tenglish) to structured English intent
↳ Fallback: Google STT v2 on Whisper service error — SLA maintained at managed-service level

Orchestration

Gemini 1.5 Flash · LangGraph · Leave Agent

↳ Intent Router (Gemini Flash): classifies intent, extracts entities, routes to specialist agent
↳ Leave Agent (LangGraph state machine): queries balance, calls Policy RAG, decides, triggers audit write
↳ HITL Manager: composes owner brief, sends WhatsApp escalation, waits for resolution webhook

Policy Layer

pgvector (Supabase) · Gemini Flash Embeddings

↳ HR Policy PDF ingested: chunked (256 tokens), embedded via Gemini Flash, stored in Supabase pgvector
↳ On policy update: owner uploads new PDF → re-indexing triggered automatically → behaviour updates immediately
↳ Every query returns: answer, source clause, page reference, confidence score

Data Layer

Firestore · Cloud Functions · Pub/Sub

↳ Firestore: employee records, leave ledger, HITL queue, session state, append-only audit log
↳ Cloud Functions: notification dispatcher, audit log writer, owner HITL webhook handler
↳ Pub/Sub: event bus decoupling all services — inbound messages, STT results, agent decisions, audit triggers

Phase C2 · Data Architecture

Entity	Store	Key fields	Retention
Employee	Firestore /employees/{id}	name, mobile, language_pref, dept, joining_date, leave_balance {casual, sick, earned}, esic_number	Duration + 3 years
Leave Request	Firestore /leave_requests/{id}	employee_id, type, dates[], status, policy_clause, rag_confidence, decision_by (AI\|HUMAN), timestamp	5 years (labour law)
Policy Document	Supabase pgvector	chunk_text, embedding[1536], source_doc_version, clause_id, page_number	All versions retained
Audit Log	Firestore /audit_log/{id}	timestamp, actor, action, entity_id, policy_clause_ids[], rag_confidence, outcome, escalated(bool)	7 years (tax compliance)
HITL Queue	Firestore /hitl_queue/{id}	request_id, context_json, ai_recommendation, owner_notified_at, resolved_at, resolution, timeout_at	1 year post-resolution
Conversation Session	Firestore /sessions/{id}	channel, employee_id, state_machine_node, context, created_at, ttl: 24h	24 hours (TTL auto-delete)

TOGAF ADM — Phase D

Technology Architecture.

Phase D maps every application component to a named infrastructure service or OSS technology. Every component has a justification, an alternative that was considered, and a reason it was rejected. This is the layer that makes the architecture buildable.

Compute · GCP

Cloud Run (asia-south1)

Scale-to-zero serverless. All agents, STT container, translation container, notification service. Free tier: 180K vCPU-seconds/month absorbs the majority of SMB workload. P99 response time for warm instances: < 3 seconds.

Scale-to-zero99.95% SLAGCP managed

LLM · GCP Vertex AI

Gemini 1.5 Flash

Intent routing and agent reasoning. Retained as managed GCP service for predictable latency SLA. At $0.075/1M input tokens — cheapest capable hosted LLM. OSS alternative (Llama 3.1 8B) rejected: P99 CPU latency exceeds 8s for complex policy queries.

$0.075/1M tokensVertex SLA

STT · OSS on Cloud Run

Whisper large-v3

OpenAI OSS model containerised on Cloud Run GPU spot. WER < 8% across 99 languages including Hindi, Telugu, Marathi. 83% cheaper than Google STT v2 at this interaction volume. Fallback to Google STT on error. Min-instance warm: eliminates cold start.

OSSGPU spot83% cost saving

Translation · OSS

NLLB-200 (Meta AI)

200-language translation model purpose-built for low-resource languages. Superior to Google Translate for Indic language code-switching (Hinglish, Tenglish). Bundled in Cloud Run container — zero marginal cost. Replaces Cloud Translation API entirely.

OSS200 languages~100% saving

Vector Store · OSS

pgvector on Supabase

Postgres + pgvector extension for HR Policy PDF RAG. Supabase free tier: 500MB database, adequate for any SMB policy document (~2MB). Equivalent retrieval quality to Vertex AI RAG Engine at SMB scale. ~60% cost reduction vs managed alternative.

OSSFree tierSupabase managed

Database · GCP

Firestore

Messaging · GCP

Pub/Sub

Firestore: serverless NoSQL. Free tier covers 50K reads/day — 83× our estimated daily volume. Append-only security rules on audit log. Pub/Sub: native GCP event fabric. Guaranteed delivery. First 10GB/month free — covers SMB entirely.

Free tierGCP nativeServerless

Voice Gateway · Regional

Exotel (India) / Plivo

India-local SIP with Mumbai and Chennai data centres. ~₹0.35/min vs Twilio's ~₹1.10/min — 68% cheaper. Lower latency via local PSTN. Provider-agnostic IVR layer: substituting with regional operator requires only gateway config change, no application code change.

India-local68% saving vs Twilio

Channel · Meta

WhatsApp Business API

Meta Cloud API — user-initiated service messages carry zero Meta fee from Nov 2024. Utility templates within 24h service window: free from Jul 2025. Business-initiated HITL alerts: ~$0.011/message in India. At < 50 escalations/month, WhatsApp cost rounds to negligible.

Free for service msgsMeta managed

Security · GCP

Secret Manager · IAM · Firestore Rules

Zero credentials in source code or environment variables. Per-service IAM identities (minimum permissions). Firestore security rules deny update/delete on audit collection. WhatsApp sender number as employee identity — verified by Meta. No passwords, no tokens for workers.

Zero-credential codeAppend-only audit

TOGAF ADM — Phase E · Migration Planning

From paper register
to autonomous system.

Migration is not a technical challenge. It is a change management challenge. The workforce has never interacted with a formal HR system. The transition must be invisible — the system must feel like an improvement on WhatsApp, not a replacement of it.

Migration sequence — Rathi Textiles deployment

Week 1–2

Owner onboarding. Priya creates the HR Policy PDF using the provided template. System indexes the document. WhatsApp Business number activated. Owner reviews HITL interface on mobile browser.

Pre-go-live

Week 3

Employee registration. Each worker sends one WhatsApp message to the system number. Their mobile number is registered as their identity. Mohammed Irfan (supervisor) acts as floor champion — introduces the system to weaving unit staff.

Go-live prep

Week 4

Soft launch — leave queries only. Workers use the system to check leave balances only. No leave applications processed yet. Builds familiarity with the interaction pattern. Priya monitors HITL queue — should be empty at this stage.

Phase 1A

Week 5–8

Full Phase 1 live. Leave applications open. All decisions processed by the system. HITL fires on < 3% of interactions. Priya receives weekly summary on WhatsApp. Old paper register retired. Labour notice risk eliminated.

Phase 1 live

Q3 2026

Phase 2 — Onboarding agent. Document collection via WhatsApp replaces paper forms. ESIC/EPF registration trigger automated. New joinee experience: entire onboarding completed on their phone before day one.

Phase 2

Q3–Q4 2026

Phase 3 — Payroll automation. WhatsApp check-in/out replaces attendance register. Payroll calculation automated. Priya's Sunday evening is freed. Payslips delivered to all 52 workers on WhatsApp — first time any of them have received a payslip.

Phase 3

Migration risk — and why it is low

The primary migration risk is worker non-adoption. This risk is structurally mitigated by the channel choice: workers are not adopting a new platform. They are using WhatsApp — which they already open 24 times a day — to interact with a new number. The behavioural change is minimal. The system does not ask workers to change how they communicate; it inserts itself into communication they were already having.

TOGAF ADM.Phases A through D.

TOGAF ADM.
Phases A through D.