Architecture Portfolio · 2026

The HR system that works the way your workforce does.

Two billion deskless workers have no HR system. Not because the technology doesn't exist — because every HR system was built for the person sitting at a desk, not the one standing at a loom.

The Autonomous HR runs on WhatsApp and a phone call. No app. No portal. No HR department required.

WhatsApp-native Voice · IVR Policy-governed RAG TOGAF ADM GCP + OSS hybrid HITL-gated Multi-language STT Deskless workforce

2.7B

Deskless workers globally Without access to a single HR system designed for them

< 60s

Leave approved, confirmed, recorded From voice note to WhatsApp confirmation

$0.15

Per employee per month Full HR automation at near-zero infrastructure cost

New apps to install Works on any Android with WhatsApp already installed

01 — Why Now

Four conditions finally aligned.

The deskless workforce has existed for all of human history. The gap in HR tooling has existed for decades. What changed is not the problem — it's the infrastructure available to solve it.

Force — 01

WhatsApp reached ubiquity in the global workforce

A factory worker in Lagos, a construction crew in Jakarta, a garment worker in Dhaka — they all share one thing: WhatsApp. Not an enterprise app. Not a portal. WhatsApp. 2.5 billion active users, the majority of them not sitting at a desk.

Force — 02

Multilingual speech recognition crossed the accuracy threshold

Whisper large-v3 achieves word error rates below 8% across 99 languages — including regional dialects, accented speech, and code-switching. For the first time, a system can hear a worker speak naturally and understand them reliably.

Force — 03

LLMs can reason against a policy document in real time

A business owner uploads a PDF. The system indexes it. Every HR decision — leave approval, grievance routing, entitlement calculation — is now governed by that document in real time. Policy changes take effect immediately, without a code deploy.

Force — 04

Serverless infrastructure collapsed the cost floor

Cloud Run, Firestore, and Pub/Sub scale to zero when idle. A 50-person business running 500 HR interactions a month pays for 500 interactions — not for a server running 24/7. The economics of enterprise HR infrastructure are now available at street-stall scale.

02 — Philosophy

Three principles.
Non-negotiable.

These aren't design preferences. They are architectural constraints derived from the reality of the workforce this system serves. Every component satisfies all three.

The channel meets the worker. Not the other way around.

No new app. No new login. No instruction manual. Every employee interaction happens on WhatsApp or a voice call — the channels already embedded in their daily lives before they ever heard of this system. An HR system that requires onboarding has already failed the workforce it claims to serve.

In practice: A worker sends a WhatsApp voice note in their language. Within 60 seconds they receive a written confirmation. They never interacted with software.

The policy document is the system. Code is just the executor.

All decision logic derives from a single HR Policy PDF uploaded by the employer. No business rules are hardcoded in the application layer. When the policy changes — a new leave quota, a revised notice period — the employer uploads a new PDF. The system's behaviour changes immediately, without a developer, without a ticket, without a deployment.

In practice: The RAG store indexes every clause. Every decision traces to a source page and paragraph. The audit log cites the policy, not the code.

III

Human oversight is a designed mechanism, not a fallback.

The system does not attempt to automate everything. When policy is ambiguous, when a legal threshold is approached, when confidence falls below 0.8 — the system pauses, composes a complete brief, and routes to the employer via WhatsApp. The employer taps Approve or Deny. This is not a failure mode. It is a first-class architectural state.

In practice: HITL triggers on <3% of interactions. Each escalation arrives as a WhatsApp message with full context, AI recommendation, and one-tap resolution.

03 — Architecture

GCP as the backbone.
OSS at the inference edges.

The customer already has GCP. The architecture keeps GCP for everything it does best — managed identity, eventing, serverless compute, and LLM inference — and surgically replaces the expensive AI inference layers with open-source equivalents where latency and accuracy guarantees hold.

System architecture — 30,000 ft view · four-layer design

The Autonomous HR — System Architecture Diagram showing four layers: Employee touch points, Channel ingestion, AI orchestration core, and Data & policy layer

Employee Channels

WhatsApp Business API · IVR Voice Gateway

↳ Meta WhatsApp Business API (Cloud) · 1,000 conversations/month free tier
↳ Exotel / Plivo SIP · local PSTN · ~$0.005/min · India / SEA / Africa DCs

Speech & Language

Whisper large-v3 · IndicTrans2 / NLLB-200

↳ Whisper on Cloud Run GPU spot · 83% cheaper than managed STT · equal accuracy
↳ NLLB-200 (Meta OSS) · 200 languages · purpose-built for low-resource languages

AI Orchestration

Gemini 1.5 Flash · LangGraph · pgvector RAG

↳ Gemini Flash on Vertex AI · intent routing + agent reasoning · $0.075/1M tokens
↳ LangGraph state machines · Leave · Payroll · Onboarding · Grievance agents
↳ pgvector on Supabase · HR Policy PDF indexed · free tier covers any SMB

Data & Audit

Firestore · Pub/Sub · Cloud Functions

↳ Firestore · employee records · leave ledger · HITL queue · session state
↳ Append-only audit log · every decision cites policy clause · 7-year retention
↳ Pub/Sub event fabric · decoupled · guaranteed delivery · $0 at SMB scale

Component	Decision	Cost saving	SLA risk
STT	Whisper OSS on Cloud Run OSS	83%	Low — fallback to managed
Translation	NLLB-200 / IndicTrans2 OSS	~100%	Low — better low-resource
LLM reasoning	Gemini 1.5 Flash GCP	Baseline	None — managed SLA
Vector store	pgvector on Supabase OSS	~60%	Low — free tier adequate
Compute	Cloud Run scale-to-zero GCP	90% vs GKE idle	None — 99.95% SLA

Design constraint

Zero idle cost. Always.

A 50-person cleaning company does not generate HR interactions at 3am. Every service scales to zero. Cloud Run, Cloud Functions, Pub/Sub — all pay-per-request. The cost model is aligned with actual usage, not theoretical peak capacity.

Security model

Mobile number as identity

WhatsApp sender number verified by Meta. IVR caller ID verified by gateway. No password, no token, no friction.

Append-only audit log

Firestore security rules deny update and delete on the audit collection. No service account holds write-over permissions.

Per-service IAM identity

Each Cloud Run service has a dedicated Service Account with minimum permissions. Zero credentials in code or environment variables.

TOGAF ADM Phases A–E → Agent Architecture → Platform & Infrastructure →

04 — The Workflow

Leave applied. Approved.
Confirmed. In 60 seconds.

A worker sends a voice note. The system transcribes it, detects the language, classifies the intent, checks the policy, validates the balance, makes the decision, writes the record, and sends the confirmation — without a single human in the loop.

Worker sends WhatsApp voice note or calls IVR

Speaks naturally in any language. No script, no menu, no keywords. "I need tomorrow off" in any of 200 languages.

Channel: WhatsApp or IVR

Whisper transcribes. NLLB-200 normalises intent.

Audio → transcript → structured intent. Language detected. Code-switching handled. Confidence scored.

< 4 seconds

Gemini Flash routes to the Leave Agent

Intent classified. Employee record retrieved from Firestore. Leave balance checked. Team headcount for requested date verified.

< 2 seconds

Policy RAG consulted

Leave Agent queries the HR Policy PDF. Governing clause retrieved with confidence score. All conditions evaluated against policy.

< 2 seconds

Decision made autonomously

If confidence > 0.8 and policy satisfied: auto-approve or auto-deny with reason. Firestore record created atomically.

< 1 second

Audit log written before notification

Immutable record: timestamp, decision, policy clause cited, confidence score, employee ID. Written before any outbound message is sent.

Synchronous write

Confirmation sent in worker's language

WhatsApp message in the detected language. Approval or denial with reason. Remaining balance included.

< 2 seconds

Total end-to-end

From voice note received to WhatsApp confirmation delivered. No human involved. No HR office needed.

< 60 seconds

HITL Path — <3% of interactions

When confidence falls below 0.8, a legal threshold is approached, or leave balance is exhausted — the system pauses. It composes a complete brief: the request, the employee record, the policy clause, the AI recommendation. The employer receives a WhatsApp message with one-tap Approve/Deny. This is not a failure mode. It is a first-class architectural state with its own audit record.

Triggers on:
Confidence < 0.8
Balance = 0
Legal threshold
Scheduling conflict

Try the Workflow Simulator → MLOps & Model Decisions →

05 — Cost Model

Enterprise HR automation.
At street-stall economics.

50 employees. 500 interactions per month. Full leave management, policy governance, multilingual voice, audit trail, HITL escalation. Running on GCP + strategic OSS. The total monthly infrastructure cost is less than a single hour of a contract HR assistant.

Component	Volume	Monthly cost
Gemini 1.5 Flash	~500K tokens/month	$0.04
Whisper STT (Cloud Run GPU spot)	50 calls × 45s avg	$0.02
Cloud Run compute	~200 invocations, warm instance	$5.20
Firestore	~20K reads / 5K writes	$0.00
Cloud Pub/Sub	~5,000 messages	$0.00
Cloud Functions	~1,000 invocations	$0.00
Voice gateway (SIP)	50 calls × 45s avg	$1.30
WhatsApp Business API	500 conversations	$0.00
Supabase pgvector (RAG store)	~500 policy chunks	$0.00
NLLB-200 translation (bundled)	~300 translations	$0.00
Total monthly — 50 employees		~$6.56

Per employee · per month

$0.13

Full HR automation. Every interaction. Every language.

vs. market alternatives (50 employees)

The Autonomous HR $6.56 / month

Typical SMB HR SaaS $250–500 / month

Mid-market HR platform $800–1,200 / month

Part-time HR assistant $400–800 / month

Enterprise HR suite $2,000+ / month

Full Cost Methodology → The Problem in Depth →

06 — Roadmap

Five phases.
One coherent system.

Leave management is the highest-frequency HR interaction and the sharpest test of the voice + policy + autonomous decision stack. It ships first. Every subsequent phase inherits the same channel, the same policy engine, and the same audit architecture.

Phase 01 — Now

Leave Management

Voice + WhatsApp channel
Multilingual STT · 200 languages
Policy RAG engine
Autonomous approve / deny
HITL escalation path
Immutable audit log

Q2 2026

Phase 02

Employee Onboarding

Document collection via WhatsApp
ID verification workflow
Bank detail capture
Digital offer letter delivery
Statutory registration trigger

Q3 2026

Phase 03

Attendance & Payroll

WhatsApp check-in / check-out
Attendance ledger automation
Payroll calculation engine
Payslip as WhatsApp PDF
Payment disbursement trigger

Q3 2026

Phase 04

Grievance Management

Voice / text grievance logging
Statutory register maintained
Escalation workflow
Resolution tracking
Compliance export package

Q4 2026

Phase 05

Predictive Intelligence

Attrition risk scoring
Leave pattern anomaly detection
Workforce demand forecasting
Employer insight dashboard
Explainable ML · SHAP

Q1 2027

Client Brief — Rathi Textiles → Architecture Decision Records →

The HR system that works the way your workforce does.

Four conditions finally aligned.

Three principles.Non-negotiable.

GCP as the backbone.OSS at the inference edges.

Leave applied. Approved.Confirmed. In 60 seconds.

Enterprise HR automation.At street-stall economics.

Five phases.One coherent system.

Three principles.
Non-negotiable.

GCP as the backbone.
OSS at the inference edges.

Leave applied. Approved.
Confirmed. In 60 seconds.

Enterprise HR automation.
At street-stall economics.

Five phases.
One coherent system.