-blue){kind=icon}
Every software startup that reaches ₹50L ARR hits the same wall — context evaporation.
| Pain | Before Sarthi |
|---|---|
| Anomaly detection | 3 weeks (if ever) |
| Runway accuracy | Monthly manual calc |
| BI query time | 2–4 hrs (analyst) |
| Context on alerts | None |
| Weekly digest | Manual assembly |
The gap: No system watches your data continuously, reasons about anomalies with memory of the past, answers natural language questions, and gets smarter with every event.
Sarthi is an ops memory layer with two focused agents:
┌──────────────────────────────────────────────────────────────┐
│ FINANCE AGENT │
│ Watches: Razorpay, bank feed, expenses │
│ Does: Burn/runway, anomaly detection, spend memory │
│ Output: Telegram alerts + weekly digest │
├──────────────────────────────────────────────────────────────┤
│ BI AGENT │
│ Watches: PostgreSQL, Sheets, GA4, Mixpanel │
│ Does: NL → SQL → chart → narrative, proactive insights │
│ Output: Charts + narrative + Telegram digest │
└──────────────────────────────────────────────────────────────┘
Value delivered:
- Anomaly detection: < 5 minutes (was 3 weeks)
- Runway accuracy: Real-time (was monthly)
- BI query time: < 30 seconds (was 2–4 hrs)
- Cost: ₹9,999/month (was ₹50,000+ for human)
┌─────────────────────────────────────────────────────────────┐
│ EXTERNAL TRIGGERS │
│ Razorpay ──┐ │
│ Bank Feed ─┼──→ Go Fiber API ──→ Redpanda ──→ Temporal │
│ Telegram ─┘ (HMAC Validated) (Event Bus) (Workflows) │
└─────────────────────────────────────────────────────────────┘
│
┌──────────────────────┘
▼
┌─────────────────────────────────────────────────────────────┐
│ PYTHON AI WORKER │
│ │
│ Temporal Activity: RunLangGraphAgent(agent, event) │
│ │ │
│ ┌────────┴────────┐ │
│ ▼ ▼ │
│ FinanceAgent BIAgent │
│ (LangGraph) (LangGraph) │
│ │ │ │
│ PostgreSQL PostgreSQL │
│ Qdrant Qdrant │
└─────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────┐
│ OUTPUT LAYER │
│ Telegram ← alerts, charts, digests, HITL buttons │
│ Langfuse ← all LLM traces, scores, costs │
└─────────────────────────────────────────────────────────────┘
| Layer | Technology | Why |
|---|---|---|
| API Gateway | Go + Fiber | High concurrency, low latency |
| Event Bus | Redpanda | Kafka-compatible, persistent |
| Workflow Engine | Temporal | Durable execution, HITL signals |
| Agent Framework | LangGraph (Python) | ReAct graphs, state machines |
| LLM | Ollama (qwen3:0.6b) | Local, no API keys, fast |
| Prompt Compiler | DSPy | Systematic, not hand-tuned |
| Memory | Qdrant | Semantic + episodic search |
| Primary DB | PostgreSQL + sqlc | Type-safe queries |
| Observability | Langfuse | LLM trace + eval scoring |
| Notifications | Telegram Bot API | Zero-friction HITL |
| Charts | Plotly Python | Code-executed, shareable PNG |
| Deploy | Docker Compose | Local dev + Hetzner |
Purpose: Continuously monitors financial events. Detects anomalies. Tracks burn and runway. Remembers context. Alerts with explanation.
LangGraph Nodes (9-node ReAct loop):
1. INGEST_EVENT → Validate, normalize, classify
2. UPDATE_SNAPSHOT → Query PostgreSQL: revenue, expense, burn, runway
3. LOAD_VENDOR_BASELINE → IF expense: query 90d avg for vendor
4. DETECT_ANOMALY → Score-based rules (spike, first-time, runway)
5. QUERY_MEMORY → Qdrant: "similar anomaly for {vendor}"
6. REASON_AND_EXPLAIN → LLM (DSPy): 1–3 sentence explanation
7. DECIDE_ACTION → ALERT (>0.7) | DIGEST (weekly) | SKIP
8. WRITE_MEMORY → Qdrant: event + explanation + outcome
9. EMIT_OUTPUT → Telegram alert or weekly digest
HITL Flow:
Founder taps [Investigate]
→ BI Agent triggered: "Break down {vendor} costs"
→ Chart + breakdown → Telegram
Founder taps [Dismiss]
→ Qdrant: "dismissed — not anomalous"
→ Future threshold raised for this vendor
Purpose: Answers natural language questions. Proactively surfaces insights. Executes SQL safely. Remembers past queries.
LangGraph Nodes (9-node ReAct loop):
1. UNDERSTAND_QUERY → Classify: aggregation/trend/breakdown
2. SELECT_DATASOURCE → Revenue → payments, Users → users
3. GENERATE_SQL → LLM (DSPy): NL → SQL
4. EXECUTE_SQL → PostgreSQL (read-only), retry on error
5. DECIDE_VISUALIZATION → line/bar/pie/text based on type
6. GENERATE_CODE → Plotly Python in sandbox
7. GENERATE_NARRATIVE → LLM (DSPy): 2–4 sentence plain English
8. WRITE_MEMORY → Qdrant: query + SQL + narrative
9. EMIT_OUTPUT → Telegram with chart + narrative
Proactive Monday Queries:
- "How did revenue change this week vs last week?"
- "What are the top 3 expense categories this month?"
- "Which user cohort has the lowest 7-day retention?"
# Docker installed
docker --version
# Ollama installed and running
ollama list
# Should show: qwen3:0.6b, nomic-embed-text, ibm/granite-docling# Clone repo
git clone https://github.com/Aparnap2/IterateSwarm.git
cd iterate_swarm
# Start Docker containers (one at a time)
docker run -d --name sarthi-postgres \
-e POSTGRES_USER=sarthi -e POSTGRES_PASSWORD=sarthi \
-e POSTGRES_DB=sarthi -p 5432:5432 \
postgres:15-alpine
docker run -d --name sarthi-qdrant \
-p 6333:6333 \
qdrant/qdrant:latest
docker run -d --name sarthi-redpanda \
-p 19092:19092 \
redpandadata/redpanda:latest-fips \
redpanda start --overprovisioned
# Temporal (requires PostgreSQL first)
docker run -d --name sarthi-temporal \
-e DB=postgresql \
-e POSTGRES_USER=sarthi \
-e POSTGRES_PWD=sarthi \
-e POSTGRES_SEEDS=host.docker.internal \
-p 7233:7233 -p 8089:8089 \
temporalio/server:latest
# Wait for all to start
sleep 30
# Verify
docker ps# Full test suite
bash scripts/test_sarthi_v2.sh
# Finance agent tests only
cd apps/ai
uv run pytest tests/unit/test_finance_nodes.py -v
# BI agent tests only
uv run pytest tests/unit/test_bi_nodes.py -v
# E2E tests (requires all containers running)
uv run pytest tests/e2e/ -v --timeout=120# Simulate Razorpay payment (2.3x AWS spike)
TENANT_ID=test-tenant-001 bash scripts/simulate_payment.sh
# Simulate BI query
TENANT_ID=test-tenant-001 QUERY="What was MRR last month?" \
bash scripts/simulate_query.sh
# Smoke test
bash scripts/smoke_test.sh[0:00] "Sarthi is a multi-agent agentic AI system — the
ops memory brain for software startups."
[0:20] Run: ./scripts/simulate_payment.sh
"Just fired a fake Razorpay webhook —
AWS bill 2.3x higher than the 90-day baseline."
[0:35] Open Temporal UI → FinanceWorkflow RUNNING
"Temporal ensures this survives any crash.
Durable execution — not a cron job."
[0:50] "LangGraph ReAct loop: Ingest → Load baseline
→ Detect anomaly → Query Qdrant → Reason → Alert"
[1:10] Show Qdrant returning memory:
"Similar AWS spike. October 2025.
Cause: undeleted staging environment."
"It didn't just detect it — it remembered."
[1:30] Show Telegram alert:
"AWS bill 2.3x usual. First spike since October.
Check recent deployments. [Investigate][Dismiss]"
[1:50] Tap [Investigate]
"Temporal receives the signal. BI Agent activates.
Generates SQL, executes it, builds a chart."
Show chart arriving in Telegram (< 30 seconds)
[2:20] Open Langfuse:
"Every LLM call traced: input, output, tokens,
latency, score. Production observability."
[2:45] "Two agents. Nine technologies.
Temporal durable workflows. LangGraph ReAct.
Qdrant episodic memory. Deployed. Tested.
Observable. This is Sarthi."
[3:00] END
| Category | Tests | Target | Status |
|---|---|---|---|
| Finance Unit | 14 | 14+ | 🔲 Pending |
| BI Unit | 10 | 10+ | 🔲 Pending |
| Go Webhooks | 5 | 5+ | 🔲 Pending |
| E2E Flows | 8 | 8+ | 🔲 Pending |
| LLM Evals | 3 | 3 | 🔲 Pending |
Run all tests:
bash scripts/test_sarthi_v2.shCurrent Phase: Week 1 — Finance Agent Implementation
| Week | Dates | Deliverable | Status |
|---|---|---|---|
| 1 | Mar 21–27 | Finance Agent end-to-end | 🔲 In Progress |
| 2 | Mar 28–Apr 3 | BI Agent end-to-end | 🔲 Pending |
| 3 | Apr 4–10 | Cross-agent + memory | 🔲 Pending |
| 4 | Apr 11–17 | Production deploy | 🔲 Pending |
| 5 | Apr 18–24 | htmx dashboard (optional) | 🔲 Pending |
| Doc | Purpose |
|---|---|
| PRD | Complete product requirements, agent specs |
| Architecture | System design, data flow |
| Testing | Testing strategy, 40+ unit + 8+ E2E |
| Vectorless RAG | ZincSearch + Docling + LangExtract |
| ZincSearch Integration | Document retrieval pipeline |
- Fork the repo
- Create a feature branch (
git checkout -b feat/your-feature) - Write tests first (RED → GREEN → REFACTOR)
- Run invariant checks before commit
- Open a PR
Invariant Checks (run before every commit):
# I-1: No raw JSON marshal in workflow/
grep -rn "json.Marshal\|json.Unmarshal" apps/core/internal/workflow/ \
| grep -v "_test.go" | grep -v "// safe:" && exit 1
# I-2: No direct AzureOpenAI() outside config/llm.py
grep -rn "AzureOpenAI(" apps/ai/src/ | grep -v "config/llm.py" && exit 1
# I-3: All existing tests still pass
cd apps/core && go test ./... -timeout=60s -q && cd -
cd apps/ai && uv run pytest tests/ -x -q --timeout=90 && cd -
# I-4: No banned jargon in agent output
grep -rn "leverage\|synergy\|utilize\|streamline" \
apps/ai/src/agents/ | grep -v "# allowed:" && exit 1MIT License — see LICENSE file.
Built with: Go 1.24 • Python 3.11 • Temporal • Redpanda • PostgreSQL • Qdrant • Ollama • LangGraph • DSPy • Langfuse
Status: 🔲 v1.0.0-alpha — In Development