Atithi OS
Atithi OS is an AI operations team for a hotel chain that never sleeps. Agents answer guests in their own language, reply to reviews, nudge room rates and dispatch housekeeping, while the General Manager approves what matters from a single inbox.
- TYPE
- Web Application
- INDUSTRY
- Hospitality
- YEAR
- 2026
- DURATION
- 8 Weeks
- PILLAR
- Product Building
What was needed.
Hotel groups run thin teams across many properties. Guest messages arrive at 2 AM, bad reviews sit unanswered, rates change by gut feel, and nobody connects the AC complaint in room 204 to the three reviews that mention it. Nivaran processes one file through fixed steps. Atithi is different: a live stream of events, a dispatcher routing each one, and specialist agents handing work to each other. It never finishes; it runs. The demo company is fictional: Pebble & Pine Hotels, 14 properties and about 1,070 rooms, from Udaipur to Mumbai.
What we shipped.
Six agents: Dispatcher: routes each event, merges duplicates, declares incidents Concierge: replies in 8 languages Reputation: drafts review replies and spots recurring issues Revenue: proposes rate changes within limits Ops: dispatches staff and tracks deadlines GM Briefing: writes the 06:30 morning brief Autonomy dial per agent and action (Off / Suggest / Ask / Auto). Guardrails are enforced in code: rate floors and ceilings, a ₹2,500 compensation cap, banned phrases, and a rule that no agent can approve its own proposal. Incident mode: when related events spike (for example, a wave of Goa flight cancellations), the agents coordinate visibly on a live graph. Go simulator: generates realistic event streams and injectable scenarios (Diwali surge, monsoon flights, viral bad review, heatwave). It only runs while someone is watching. Screens: live operations screen, approval inbox with edit-and-approve, guest chat simulator, reputation, revenue rate grid, ops board, briefings, agent run explorer, settings. Cost control: simulated events replay precomputed agent output; only visitor actions call the model, within budget caps.