Role : Agentic Orchestration Lead - Autonomous

Location : India (Bengaluru / Gurugram / remote-India)

Reports to : Group Head of Technology

Team : Founding hire. Builds the platform, then a small team of 24 strong ICs.

The mission :

Build an always-on software system that runs the back office of a global hospitality platformsensing the state of every property and every corporate workflow, deciding what needs to happen, acting within bounded authority, and driving each task to verified closure across the responsible human teams, escalating only true exceptions.

You are not building a chatbot or a dashboard. You are building an agent platform : a shared substrate on which the first agent (live hotel operations) is the reference implementation, and onto which finance, payroll, tax, safety, and every other corporate function plug in over time as policy + tools, not as new from-scratch builds.

The end state is a thin control towera small number of controllers, approvers, and exception handlersin place of large HQ processing teams. Routine work runs itself; humans own judgment, approvals, and statutory sign-off.

What you will build :

- The platform substrate (build once, every agent reuses) :

1. Identity, permissions, and the action-authority matrix

2. Integration/tool layer to the operational stack (PMS, ERP/accounting, payroll, banking, tax portals, ticketing, messaging)

3. The case + closure engine opens a task, assigns it, chases it, verifies the work actually happened, and only then closes

4. Escalation and approval routing with SLA ladders

5. Eval, audit log, and observability for non-deterministic actions

6. Per-jurisdiction policy/knowledge packs

- The first tenantlive operations heartbeat : an event-triggered + scheduled loop that, per property, continuously senses state (occupancy, OOO rooms, pricing signals, guest messages, housekeeping, maintenance, compliance) and runs it to closure across the responsible teams.

- The path to a fleet : the platform makes each subsequent agent (AP, AR, reconciliations, health & safety, security, payroll, tax) mostly config + policy + eval, shipping faster as the substrate hardens.

What you will own :

- The closed loop : sense, detect, decide, act, follow up to verified closure, escalate, log.

- The action-authority tiers (autonomous / notify-after / approve-before / human-executes-statutory) and the guardrails that keep every action inside them.

- Closure verification proving a task is actually done, not just marked done.

- Multi-team SLA orchestration and escalation ladders (property, area, regional, control tower).

- Reliability of consequential, non-deterministic actions : idempotency, dry-run, rollback.

- The eval harness and audit trail that make the system trustworthy and auditor-legible.

- Cost-per-decision economics at the scale of thousands of properties.

The hard problems (the seniority filter) :

- Closure verification. The single hardest part. How do you know the room was actually re-cleaned, the ticket actually resolved, the invoice actually matchednot just stamped done?

- Bounded autonomy. Designing an authority matrix that auto-acts on the reversible and routes the irreversible to a human, with no gaps.

- Action safety. Never double-charge, double-message, or double-file. Idempotency and rollback on everything that touches money, inventory, or a guest.

- Eval before execution. Catching a wrong or hallucinated action before it fires.

- Frontline trust. GMs and staff must lean on it, not route around it.

- Multi-jurisdiction. 35+ countries means taxation or payroll is a framework plus per-country policy packs, never a single agent.

Must-haves (proof, not potential) :

- Shipped a production system that took consequential actions autonomouslymoney, inventory, or customer-facingnot an advisory tool.

- Built a workflow/case system with SLA-driven follow-up and escalation across human teams (real-time ops automation at a delivery, logistics, fintech, QSR-chain, or marketplace company is the closest analog).

- Designed human-in-the-loop approval tiers and clean escalation handoffs.

- Built eval and observability for non-deterministic / agentic systems.

- Integrated against messy, flaky third-party APIs at scale.

Strong pluses :

- Platform thinkingbuilt shared infrastructure others built on top of.

- Multi-location physical-operations domain (hospitality, retail, logistics, food).

- Exposure to financial-controls / audit-grade systems (segregation of duties, audit trails).

- Has led a small, senior IC team.

Success looks like :

- 90 days : platform substrate stood up; operations heartbeat live on a pilot cluster doing sense + detect + a handful of autonomous actions; full audit log.

- 180 days : closed-loop on a defined operations case set (OOO, housekeeping, guest-message SLA, rate band) with verified closure and working escalation; measurable drop in human touches and SLA breaches.

- 365 days : majority of routine operations cases handled end-to-end with humans only on exceptions; second-wave agents (AP, AR, H&S, security) shipping on the same platform; eval harness mature.

What this role is NOT :

- Not building a small/domain language model yetthat is a later phase, trained on the data this platform generates.

- Not rebuilding the PMS or the guest product.

- Not a research role. This ships to production under real reliability constraints.

Disqualifiers :

- Treats it as a chatbot or a dashboard.

- Cant define done or describe how to verify closure.

- No eval discipline.

- Over-automates with no escalation or approval design