Early-stage startups drown in operational chaos:

The gap: Not "startups don't have tools" — it's 15 disconnected tools with no intelligence layer connecting them.

Sarthi is the internal operations virtual office for Seed to Series A startups.

We don't find you customers.
We make sure your company doesn't collapse while you do.

Every operational task that doesn't require your
unique human judgment — Sarthi handles.
Everything that does — Sarthi prepares perfectly
and puts in front of you in 30 seconds, not 3 hours.

"Sarthi saved me 18 hours this week:

• Prepared my GST + VAT filing data automatically (India + UK)
• Sent payment reminders to 5 overdue clients
• Told me my runway dropped from 9 to 7 months and showed me exactly why
• Drafted the offer letter for my new UK hire (compliant with UK law)
• Found ₹23k/month in unused SaaS subscriptions
• Tracked all contract renewals and flagged 2 expiring soon"

That's not a feature list. That's a Wednesday.

❌ RevOps / GTM / CRM outreach ❌ Customer success / support ❌ External market intelligence (competitors, pricing) ❌ Content generation / marketing ❌ Cap table management ❌ Tax filing ❌ Grant applications

Why? Internal ops is painful enough. We solve that completely.

┌─────────────────────────────────────────────────────────────────┐
│  TIER 0 — KERNEL (Go + Temporal + Graphiti)                     │
│  BusinessOSWorkflow: orchestrates all agents, manages state     │
│  enforces HITL gates, temporal knowledge graph                  │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│  TIER 1 — CHIEF OF STAFF (1 agent)                              │
│  The only agent that talks to the founder.                      │
│  Routes work to 6 desks.                                        │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│  TIER 2 — 6 DESKS (13 Virtual Employees)                        │
│                                                                  │
│  ┌─────────────┐ ┌─────────────┐ ┌─────────────┐               │
│  │ Finance     │ │ People      │ │ Legal       │               │
│  │ Desk (4)    │ │ Desk (2)    │ │ Desk (2)    │               │
│  └─────────────┘ └─────────────┘ └─────────────┘               │
│                                                                  │
│  ┌─────────────┐ ┌─────────────┐ ┌─────────────┐               │
│  │ Intelligence│ │ IT & Tools  │ │ Admin       │               │
│  │ Desk (2)    │ │ Desk (1)    │ │ Desk (2)    │               │
│  └─────────────┘ └─────────────┘ └─────────────┘               │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│  TIER 3 — DATA LAYER (ingest + memory, never surfaces)          │
│  Ingestion | Memory (Qdrant + Neo4j) | Connector                │
└─────────────────────────────────────────────────────────────────┘

Total monthly cost for MVP: $0

Role: The face of Sarthi. Routes work to 6 desks, synthesizes intelligence, manages the relationship.

Output:

• "Here's what I found + one action"
• "I handled X — here's what I did"
• "I need your decision on X — here's context"
• "Weekly briefing: X handled, Y needs you"

CFO Agent

• 13-week rolling cash flow forecast
• Burn rate + runway calculation
• Unit economics: CAC, LTV, payback
• Scenario modeling: "what if we hire in March?"

Bookkeeper Agent

• Expense categorization (auto)
• Bank vs books reconciliation
• Monthly P&L narrative generation

AR/AP Clerk Agent

• Invoice generation + payment reminders
• Accounts receivable aging report

Payroll Clerk Agent

• Payroll data preparation
• PF/ESIC/pension filing reminders

HR Coordinator Agent

• Onboarding checklist execution
• Offer letter generation
• Leave balance tracking

Internal Recruiter Agent

• Job description drafting
• Interview scheduling coordination

Contracts Coordinator Agent

• NDA generation (templates)
• Contract review summary
• Contract expiry tracking

Compliance Tracker Agent

• GST, TDS, advance tax deadlines
• PF, ESIC, PT compliance
• DPDP Act / GDPR compliance checklist

BI Analyst Agent (Internal-Only)

• Customer cohort analysis (retention, churn)
• Revenue concentration risk
• Anomaly detection across all data sources

Policy Watcher Agent

• Regulatory change monitoring
• Tax law updates affecting THIS company

IT Admin Agent

• SaaS subscription audit + optimization
• Cloud spend analysis
• Tool access provisioning/deprovisioning

Executive Assistant Agent

• Meeting prep (pulls context, agenda)
• Action item extraction from meeting notes

Knowledge Manager Agent

• SOP documentation from observed workflows
• Internal wiki organization

• IngestionAgent: Normalizes all data (bank PDF, accounting, Notion, Slack, Stripe)
• MemoryAgent: Qdrant + Neo4j + Graphiti — company's long-term brain (vector + temporal)
• ConnectorAgent: OAuth + API connections to all tools

How Sarthi gets smarter about YOUR company over time:

1. AGENT ACTS → Bookkeeper categorizes AWS expense
2. OUTCOME OBSERVED → Founder confirms category is correct
3. MEMORY UPDATED → "AWS expenses: categorized as 'Cloud Infrastructure'" (Qdrant + Neo4j)
4. FUTURE ADJUSTED → Auto-categorize similar expenses, flag anomalies
5. CONTEXT DRIFT DETECTED → "Cloud spend up 40% MoM. New deployment or pricing change?"

Not generic AI. Company-specific intelligence.

Replace ₹2.0L–₹3.75L/month in admin costs with ₹5K–₹15K/month.

• Docker & Docker Compose
• Azure OpenAI account (or OpenRouter)
• Telegram Bot Token (from @BotFather, free)

# Clone
git clone https://github.com/Aparnap2/IterateSwarm.git
cd iterate_swarm

# Start infrastructure
docker compose up -d

# Wait for services
sleep 60

# Check health
docker compose ps

1. Message your Telegram bot: /start
2. Complete 6-question onboarding
3. Upload first bank statement (CSV/Excel/PDF)
4. Receive first insight: "Here's your cash position + one action"

# All tests
cd apps/ai
uv run pytest tests/ -v --timeout=120

# Specific groups
uv run pytest tests/test_sarthi_tdd.py -v           # TDD tests (~80)
uv run pytest tests/test_e2e_saarathi.py -v         # E2E flows (20)
uv run pytest tests/test_llm_eval.py -v             # LLM evals (15)
uv run pytest tests/test_llm_connectivity.py -v     # LLM connectivity (4)

# Phase 3 Tests (TBD - Not Yet Available)
# uv run pytest tests/test_finance_desk.py -v         # Finance Desk (25) - Phase 3
# uv run pytest tests/test_people_desk.py -v          # People Desk (12) - Phase 3
# uv run pytest tests/test_legal_desk.py -v           # Legal Desk (12) - Phase 3
# uv run pytest tests/test_intelligence_desk.py -v    # Intelligence Desk (15) - Phase 3
# uv run pytest tests/test_it_desk.py -v              # IT Desk (10) - Phase 3
# uv run pytest tests/test_admin_desk.py -v           # Admin Desk (12) - Phase 3

# With coverage
uv run pytest tests/ --cov=src --cov-report=html

Test Categories:

1. Infrastructure Health (6 tests) — Azure LLM, Qdrant, Neo4j, PostgreSQL reachable
2. Memory Agent (22 tests) — Embeddings, Qdrant upsert, Neo4j graph, semantic search
3. Chief of Staff (8 tests) — Plain language, correct routing, ToneFilter fidelity
4. Bank Parser (8 tests) — HDFC/ICICI/SBI CSV, Docling accurate mode, multi-format
5. Finance Desk (25 tests) — CFO, Bookkeeper, AR/AP, Payroll agents - Phase 3 TBD
6. People Desk (12 tests) — HR Coordinator, Internal Recruiter - Phase 3 TBD
7. Legal Desk (12 tests) — Contracts, Compliance Tracker - Phase 3 TBD
8. Intelligence Desk (15 tests) — BI Analyst, Policy Watcher
9. IT & Tools Desk (10 tests) — IT Admin
10. Admin Desk (12 tests) — EA, Knowledge Manager
11. E2E Flows (20 tests) — Full stack: onboarding, bank ingestion, HITL approval
12. LLM Evals (15 evals) — LLM-as-judge for tone, jargon, actionability

Test Status Target: ~189 tests passing (165 current + 24 Phase 3 TBD)

v4.2.0 Status: v4.2.0-alpha — Internal Ops Virtual Office Only.

See docs/TESTING_ARCHITECTURE.md for complete testing docs.

1. Create apps/ai/src/agents/desks/your_agent.py
2. Define role, tools, APIs, output format (Pydantic)
3. Register in Chief of Staff routing
4. Add tests in apps/ai/tests/test_your_agent.py
5. Run: uv run pytest tests/test_your_agent.py -v

# LLM (swap provider by changing these 3 lines)
LLM_BASE_URL=https://openrouter.ai/api/v1
LLM_API_KEY=sk-or-v1-...
LLM_MODEL=openai/gpt-4o-mini

# Azure OpenAI (Alternative)
AZURE_OPENAI_API_KEY=your-key
AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com/
AZURE_OPENAI_DEPLOYMENT_NAME=gpt-4o-mini
AZURE_OPENAI_API_VERSION=2024-02-01
# Set these instead of LLM_BASE_URL/LLM_API_KEY/LLM_MODEL when using Azure.

# Telegram
TELEGRAM_BOT_TOKEN=...

# Neo4j / Graphiti
NEO4J_URI=bolt://localhost:7687
NEO4J_USER=neo4j
NEO4J_PASSWORD=...

# PostgreSQL
POSTGRES_HOST=localhost
POSTGRES_PORT=5432
POSTGRES_DB=sarthi

# Qdrant
QDRANT_HOST=localhost
QDRANT_PORT=6333

# Sandbox
SANDBOX_URL=http://localhost:5001
SANDBOX_SECRET=...

• LLM unification (get_llm_client everywhere)
• Graphiti + Neo4j full integration
• 125 tests passing

• Finance Desk (CFO + Bookkeeper + AR/AP + Payroll)
• People Desk (HR + Internal Recruiter)
• Legal Desk (Contracts + Compliance)
• 150 tests passing

• Intelligence Desk (BI + Policy Watcher)
• IT & Tools Desk (IT Admin)
• Admin Desk (EA + Knowledge Manager)
• 175 tests passing

• Chief of Staff routing (internal-only)
• BusinessOS workflow (Go + Temporal)
• HITL gate E2E test
• 20/20 E2E tests green

• One real founder onboards
• Uses at least 2 desks
• Reports "This saved me admin time"
• → TAG v4.2.0

1. Fork the repo
2. Create a feature branch (git checkout -b feat/your-feature)
3. Write tests first (RED → GREEN → REFACTOR)
4. Commit with conventional commits
5. Open a PR

Note: All agents must have 100% test coverage before merge.

MIT License — see LICENSE file.

Built with: Go 1.24 • Python 3.13 • Temporal • Redpanda • PostgreSQL • Qdrant • Neo4j • Graphiti • Azure OpenAI • LangGraph • DSPy • Pydantic v2

Status: ✅ v4.2.0-alpha — Internal Ops Virtual Office Only