Fourteen AI agents run the work behind an event, each paired with a human lead.
Agents propose, policy decides, humans approve, code executes.
The problem. Most college events run on WhatsApp threads, spreadsheets and favours. When a speaker cancels, the change has to reach rooms, volunteers, announcements, helpdesk answers and sponsors, and people usually update only some of them. Enterprise event tools cost lakhs, are built for corporate conferences, and do not handle Indian college needs such as faculty approval and OD letters.
The solution. OFFSTAGE gives every event a team of 14 specialised agents that share one source of truth. Agents never write to the database directly. They file proposals, and a policy engine written in code decides which ones need a human and which human. Approved proposals run through deterministic executors in one database transaction. Every step is traced, every message goes through an outbox, and emergencies always go to people.
flowchart LR
A["Agents propose<br/>14 agents file typed proposals"] --> P{"Policy decides<br/>code assigns T0 to T3"}
P -- "T0, T1" --> X["Code executes<br/>one DB transaction, audit log, outbox"]
P -- "T2: 1 approval<br/>T3: 2 approvals" --> H["Humans approve<br/>the agent's lead, owner or faculty"]
H --> X
X -- "new domain events" --> A
T1 actions run at once with a 10 minute undo window. Emergencies (medical, safety, fire, harassment) are never handled by agents: they raise a loud alert to every lead and agents only help draft. Details: docs/policy.md.
Every agent is a config on one shared runtime (src/agents/index.ts). Tiers below are the tiers the policy engine assigns to each agent's typical action; full detail in docs/agents.md.
| Agent | Job | Human lead | Example action | Tier |
|---|---|---|---|---|
| Commander | Turns a disruption into one plan; intake, daily briefing, what-if | Event head | Speaker cancels: move a session into the gap, reassign crew, announce | T2 (T3 if a 50+ session is cancelled) |
| Planner | Flags overdue and at-risk milestones | Event head | "Overdue: venue booking confirmed" note | T0 |
| Finance | Warns before overspend and proposes cover; never pays | Treasurer | Move budget from Decor to Food | T2 (T3 above 10,000 INR) |
| Sponsorship | Drafts sponsor follow-ups; never sends | Sponsorship lead | Follow-up mail draft for a sponsor | T1 |
| Marketing | Registrations against target, push suggestions, post drafts | Marketing lead | "25% behind target: push these 3 colleges" | T0 |
| Registrar | Flags duplicates, promotes the waitlist | Registrations lead | Promote 10 people from the waitlist | T2 (T3 if bulk) |
| Scheduler | Keeps the schedule clash-free with a deterministic solver | Program lead | Shift downstream sessions when one runs late | T2 |
| Speaker Liaison | Reminds speakers to send AV and travel requirements | Program lead | Requirements reminder to one speaker | T1 |
| Crew Chief | Covers missed shifts fairly with a solver | Volunteer lead | Assign Ravi to the registration desk shift | T1 |
| Logistics | Food counts and room readiness checklists | Logistics lead | "Lunch: 412 plates" food count | T1 |
| Herald | Announces approved changes to the right people | Comms lead | "Moved: Intro to ML" to 120 attendees | T2 (T3 above 200 people) |
| Helpdesk | Answers with citations, or escalates | Comms lead | Escalate a question it cannot ground | T0 |
| Radar | Watches question spikes, queues and voice-note reports | Ops lead | Lunch confusion: incident, signage task, announcement | T0 to T2 |
| Chronicler | Final report and lessons for next time | Event head | Final report for the event | T1 |
|
Live Stage. The console's home: every agent as a node, proposals and runs streaming in live over SSE. 
|
Ripple view. Before you approve, see everything a plan touches: sessions, people, crew, messages. 
|
|
Glass box. Every agent run is open: steps, tool calls, model, tokens, cost and latency. 
|
Jarvis voice. Talk to the Commander. It answers from live state and turns requests into proposals, never approvals. 
|
|
Helpdesk with citations. Attendees get answers grounded in event documents, with sources, or an honest escalation. 
|
Offline signed tickets. Ed25519-signed QR tickets verify on a volunteer's phone with no network. First scan wins. 
|
|
Confusion Radar. When many people ask the same thing in a few minutes, Radar opens an incident and proposes a fix. 
|
What-if. Ask "what if 30% more people show up" and see the impact in a sandbox that never touches real state. 
|
|
Daily briefing. Every morning at 07:00 IST: what happened, what is due, what is at risk, what waits for you. 
|
Close-out report. Attendance, helpdesk, incidents and money, with every number taken from SQL. 
|
|
Emergency path. Medical or safety reports skip the agents and raise a loud alert to every lead and the public status page. 
|
Speaker view. Every role has a screen: console, crew app, attendee portal, public page, and the speaker's own phone. 
|
flowchart TB
subgraph Clients
O["Organizer and leads<br/>/console"]
V["Volunteers<br/>/crew PWA"]
AT["Attendees<br/>/me, /e/slug"]
end
subgraph Web["Next.js 16 app"]
R["Route handlers<br/>authz + zod on every route"]
ACT["actions.propose / approve<br/>src/server/actions"]
POL["Policy engine<br/>src/server/policy"]
EX["Executors<br/>one transaction per proposal"]
SSE["SSE streams<br/>console + public status"]
end
subgraph DB["Postgres 17 + pgvector"]
T[("Domain tables, proposals,<br/>audit_log, domain_events")]
OB[("outbox")]
KB[("kb_chunks<br/>vector 384 + full text")]
Q[("pg-boss queues")]
end
subgraph Worker["Worker (Node)"]
D["Domain-event dispatcher<br/>+ agent cron jobs"]
AG["Agent runtime<br/>14 agent configs, solvers"]
DR["Outbox drain<br/>every 3 s"]
end
subgraph Router["Model router"]
G["Groq<br/>gpt-oss-120b / 20b, Prompt Guard 2, Whisper"]
B["Amazon Bedrock<br/>Nova 2 Lite"]
OL["Ollama<br/>qwen3:4b-instruct"]
RU["Rules-only fallback"]
end
subgraph Channels
WA["WhatsApp (Twilio)"]
SMS["SMS (Twilio)"]
TG["Telegram Bot API"]
EM["Email (SMTP: Mailpit or SES)"]
IA["In-app notifications"]
end
O & V & AT --> R
R --> ACT --> POL
POL -- "T0, T1 or approved" --> EX --> T
EX --> OB
T -- "LISTEN / NOTIFY" --> SSE --> O & AT
T -- "NOTIFY" --> Q --> D --> AG
AG -- "proposals" --> ACT
AG <--> Router
G -. "fallback" .-> B -. "fallback" .-> OL -. "fallback" .-> RU
AG --> KB
OB --> DR --> WA & SMS & TG & EM & IA
More diagrams (proposal flow, SSE stream, router chains): docs/architecture.md.
The live demo is a zero-cost build of the same app. How it works: docs/showcase-mode.md. How it is deployed: Deploying the showcase.
Showcase (NEXT_PUBLIC_SHOWCASE=1) |
Full version | |
|---|---|---|
| Data | Recorded from real runs of the full stack, PII scrubbed, replayed in the browser | Postgres 17 + pgvector |
| Agents and AI | Replayed agent runs, proposals, briefing, report and what-if | Live agents on Groq, Bedrock or Ollama |
| Approvals | Real role rules (T3 still needs two people), state kept in your browser | Policy engine, approvals and executors in Postgres |
| Helpdesk | BM25 search over the recorded knowledge base, with citations and escalation | Hybrid vector and full-text search with a model, guard and citations |
| Voice | Browser speech recognition and synthesis | Groq Whisper and Murf text to speech |
| Messages | None sent; delivery shown as simulated | WhatsApp, SMS, Telegram, email, in-app through the outbox |
| Sign-in | Pick a persona, no email or code | Email OTP (Better Auth) |
| Server API | Every /api route returns 404; a network guard blocks third-party hosts |
Full API with authz on every route |
| Cost | Static hosting only | Model and messaging usage |
Measured during the hackathon.
| What | Result |
|---|---|
| Helpdesk answers grounded in a cited source | 95 percent |
| Refusal when no source exists | 100 percent |
| Prompt injections blocked | 20 of 20 |
| Solver checks passed (schedule and crew) | 42 of 42 |
| Retrieval accuracy | 97.5 percent |
Full demo run (71 agent runs, AI_PROFILE=demo on Groq and Bedrock) |
about USD 0.03 (USD 0.0265) |
Tests in this repo today: 659 unit tests in 41 files (pnpm test), plus 14 database integration test files (pnpm test:db), Playwright end-to-end specs and a showcase network test.
| Layer | Choice |
|---|---|
| Web | Next.js 16 (App Router), React 19, TypeScript 5.9 strict, Tailwind CSS 4 |
| Contracts | zod 4 schemas in src/contracts, shared by app, worker and agents |
| Database | PostgreSQL 17 + pgvector, Drizzle ORM |
| Jobs | pg-boss on the same Postgres |
| Realtime | Server-Sent Events over Postgres LISTEN/NOTIFY |
| Auth | Better Auth with email OTP, one authz helper on every route |
| AI | Vercel AI SDK 7 with Groq, Amazon Bedrock and Ollama providers |
| Retrieval | bge-small-en-v1.5 embeddings (384 dims) run locally, hybrid search with reciprocal rank fusion |
| Solvers | Deterministic TypeScript for schedule and crew assignment |
| Channels | Twilio (WhatsApp, SMS), Telegram Bot API, nodemailer (Mailpit, Amazon SES SMTP) |
| Tickets | Ed25519 signed QR codes, verified offline with WebCrypto |
| Visuals | React Three Fiber, React Flow, GSAP |
| Tests | Vitest, Playwright, a custom eval harness |
| Hosting | Showcase on Vercel; full stack on Docker Compose with Caddy |
Try the showcase (nothing to install)
Open offstage-live.vercel.app, choose a persona (start with Event head), then trigger a scenario such as "Speaker cancels" from the Live Stage and approve the plan. Use Reset demo in the header to start again.
Run the showcase locally (no database, no .env, no keys)
Needs Node 24 and pnpm 11 (corepack enable).
pnpm install
NEXT_PUBLIC_SHOWCASE=1 pnpm build
pnpm start # http://localhost:3000Run the full stack (Docker, Postgres, worker, real agents)
Needs Node 24, pnpm 11 and Docker. Model keys are optional: with none, agents fall back to rules. Full guide: docs/running-locally.md.
cp .env.example .env
pnpm keys # paste the printed secrets and the local DATABASE_URL into .env
docker compose up -d db mailpit # Postgres 17 + pgvector, Mailpit inbox on :8025
pnpm install
pnpm db:migrate && pnpm db:seed
pnpm dev # terminal 1: http://localhost:3000
pnpm worker # terminal 2: agents, jobs, outboxsrc/
agents/ 14 agent configs and the shared agent runtime
ai/ model router, prompt-injection guard, RAG, voice, evals
app/ Next.js routes: (public), (attendee), (crew), (console), api
components/ UI kit and the console, crew, attendee, public and landing screens
contracts/ zod schemas and types shared by everything
db/ Drizzle schema, migrations, seed, demo reset and triggers
lib/ API client, time (UTC stored, IST shown), i18n, logger
server/ authz, policy engine, actions and executors, channels, SSE
showcase/ zero-cost demo: flag, recorded fixtures, replay engine, guards
solvers/ deterministic schedule and crew solvers
styles/ design tokens and theme
templates/ event type templates (hackathon, workshop, charity drive, other)
worker/
index.ts pg-boss worker: dispatcher, cron, briefing, outbox drain
jobs/ agent jobs, domain events, outbox
tests/
unit/ integration/ e2e/ evals/ voice/
scripts/ key generation, fixture recorder, deploy helpers
docs/ architecture, agents, policy, showcase, running locally, ADRs
 |
 |
 |
 |
- Executors for certificates and OD letters (both are already tiered T3 in the policy engine).
- An API route for the per-event agent kill switch; today it is a database flag checked on every agent wake.
- Native SES and Resend email drivers (email goes through SMTP today).
- More event templates (four ship today: hackathon, workshop, charity drive, other).
- Production WhatsApp senders in place of the Twilio sandbox.
Team MASTICODE. 1st Prize, AIML Hacks 2026, KG Reddy College of Engineering.
| Name | Role | GitHub |
|---|---|---|
| Vedant Idlgave | Agents and AI | @vedant7007 |
| Abhinav Nakka | Platform and security | @Abhinav1480 |
| V Thanishka | Design and experience | @vthanishka |
- AIML Hacks 2026 and KG Reddy College of Engineering for the problem statement (GENAI-25, Multi-Agent Event Management System) and the 36 hours.
- Groq, Amazon Bedrock and Ollama for the models; Hugging Face Transformers.js and
bge-small-en-v1.5for local embeddings. - pgvector, pg-boss, Drizzle, Better Auth, the Vercel AI SDK, Twilio and Telegram for the building blocks.
Questions, pilots or feedback: email us.
Released under the MIT License, copyright (c) 2026 Vedant Idlgave, Abhinav Nakka and V Thanishka. Contributions are welcome: see CONTRIBUTING.md, the Code of Conduct and the security policy. The codename in the source is sutradhar, "the one who holds the strings".