Find the deviation
Monitor governed metrics rather than asking a model to browse raw tables.
Agentic analytics · public reference build
MarginGuard AI
A governed revenue-investigation system that validates data before explanation, uses controlled analytical tools, links every conclusion to evidence, and keeps consequential actions behind human approval.
Public mode: verified replay. No public LLM, database or paid API call is made.
Business problemRevenue leakage
Engineering layerPostgreSQL + dbt
Agent boundaryTyped tools only
Control modelApprove + verify
01 · THE CASE
Not a chatbot. A controlled decision loop.
The demo business is Northstar Coffee Club, a synthetic subscription retailer. The system separates four questions that are often blurred together: is the data trustworthy, is the signal unusual, what caused it, and what action is safe to take?
Check evidence first
Freshness and completeness gates can stop the investigation before a false narrative forms.
Separate advice from action
Customer, pricing and campaign changes require explicit approval.
Measure what happened next
An action is not a success until the target outcome is checked.
02 · INTERACTIVE VERIFIED REPLAY
Choose a failure mode.
Each scenario replays a captured reference run. The payment-failure case comes from the verified PostgreSQL/dbt execution; the other fixtures exercise the same decision and guardrail patterns without exposing a public backend.
Portfolio-safe public modev1.0.0
CONTROLLED WORKFLOW
Observe → Validate → Investigate → Recommend → Approve → Verify
Ready. No conclusion is displayed before the evidence path runs.
Governed tool trace
Every conclusion has an inspectable path.
Select a scenario and run the investigation.
Evidence store
Evidence, not confidence theatre.
Human approval gate
Recommendation pending
The action boundary remains closed until the investigation completes.
✓
Outcome verification
03 · REFERENCE IMPLEMENTATION
The public demo is static. The repository is executable.
The browser layer is deliberately zero-cost and abuse-resistant. The source repository contains the reproducible warehouse, dbt project, typed investigation tools, audit/evidence store, deterministic orchestration, tests and local API.
01Synthetic business eventsPayments, customers and scenario fixtures
→
02PostgreSQL + dbtStaging, facts, marts and assertions
→
03Governed toolsTyped inputs, allow-listed metrics, read-only role
→
04Audit + evidenceRun IDs, tool calls and durable evidence records
→
05Approval + verificationConsequential action remains controlled
Public website
Safe interactive replay
- No OpenAI key
- No public database
- No user-generated tool calls
- No ongoing inference cost
GitHub repository
Full reference system
- Docker/PostgreSQL setup
- dbt models and tests
- FastAPI and audit trace
- Deterministic scenario runner
Design boundary
Agent does not receive SQL
- Enums instead of table names
- Bound parameters
- Physical read-only role
- Data-quality hard stop
BUILT BY SERGEY KAMINOV
Analytics engineering with an operating decision in mind.
MarginGuard connects my background in payments, fraud, customer behaviour and commercial analytics with a production-minded approach to agentic systems: governed tools, explicit failure modes, human judgement and outcome verification.