Marti Soura Vamseekar

Work / MaSoVa Enterprise Fleet

MaSoVa Enterprise Fleet

A Manager Copilot for a restaurant fleet: one conversational agent that fans out to seven ops specialists and puts every action behind human approval.

Source availableAug 2026Built for the All Things Agentic Hackathon
SVamseekar/masova-enterprise-fleet

The problem

Agentic ops demos usually assume a manager wants an autonomous brain. Real managers want the analysis done and the decision kept: they are accountable for the price change, the purchase order and the refund, so an agent that executes on its own is unusable.

Agent fleet

ManagerVoice or textasks in plain languageConductor agentroutes · compares · answersOperations manualretrieval · grounded answersSeven specialists — none executesDemandforecastInventoryreorderChurnpreventionReviewresponsesShiftrosteringKitchencoachingDynamicpricingApproval queue — nothing executes until a manager acceptsprices · purchase orders · refunds · campaignsRun log · reasoning trace · hash-chained
One conductor, seven specialists, and a gate nothing gets past. Every agent drafts; the manager decides.

A manager asks a question by voice or text. The conductor agent grounds policy answers in the operations manual using retrieval, and routes work to any of seven specialists: demand forecast, inventory reorder, churn prevention, review responses, shift rostering, kitchen coaching and dynamic pricing. Every agent emits a draft proposal rather than executing. Proposals converge on an approval queue, where nothing reaches a price, a purchase order, a refund or a customer until a manager accepts it. Each run is written to a hash-chained log with its reasoning trace.

How it works

  1. 01

    Ask in plain language, by voice or text

    One conversational front door. Gemini transcribes spoken input and can answer aloud, so the manager can ask while walking the floor.

  2. 02

    The Copilot decides who to ask

    It routes to whichever of the seven specialists the question needs — forecast, stock, churn, reviews, shifts, kitchen coaching or pricing — and can compare stores in one thread.

  3. 03

    Answers are grounded, not improvised

    Policy questions retrieve from the actual operations manual rather than the model's memory, and numbers come from tools that query the platform.

  4. 04

    Every action waits for a human

    Agents emit drafts. Prices, purchase orders, refunds and campaigns sit in an approval queue until a manager accepts them, and each run is hash-chained with its reasoning trace.

Engineering notes

  • Built on Google ADK with Gemini: a conductor agent (manager_chat_agent) that can trigger any of seven specialist agents as tools.
  • Seven ops agents in src/masova_agent/agents/ — demand forecasting, inventory reorder, churn prevention, review response, shift optimisation, kitchen coach and dynamic pricing.
  • RAG over data/knowledge/ answers operations-manual questions with retrieved context instead of model recall.
  • A shared AgentRuntime carries policy, reasoning-chain audit and rule-based fallbacks for when the model is unavailable.
  • Manager console at GET /console shows the live agent registry, run history with traces, a SHA-256 chain integrity badge and the approval queue — over real data, not fixtures.

Trade-offs

Proposals over autonomous execution

instead of letting agents act directly

The manager is accountable for the outcome. An agent that changes a price on its own moves the liability without moving the authority.

A conductor plus specialists

instead of one general agent with every tool

Narrow agents are testable and their failures are legible. A single agent holding twenty tools is neither.

Rule-based fallbacks

instead of failing when the model is down

A restaurant at rush cannot wait on an API. Deterministic fallbacks keep the operation running with reduced capability rather than none.

Evidence

8

agents incl. conductor

7

ops specialists

0

actions without approval