What it does

Siren lives in your Telegram group. Say “split €60 for dinner 4 ways” and it creates a PayPal invoice for each person, DMs them a pay link, tracks who paid via webhooks, and posts updates to the group. It doesn’t stop there. Siren remembers how each member pays — speed, reliability, preferred tone — and decides on its own when and how to nag. A reliable friend gets “no rush, €15 whenever”; a repeat late-payer gets “third reminder, €15 still owed.” The agent chooses, not a script. Try it: /start in any TG group with @siren_pay_bot · Sandbox test accounts in testing instructions below.

How we built it

PayPal Agent Toolkit (@paypal/agent-toolkit/ai-sdk) — official toolkit, pre-wired PayPal tools for invoicing. Judge criterion “thorough and skillful use of PayPal Developer Platform” met at the import line. Vercel AI SDK — generateText({ tools, maxSteps: 6 }) is the agent loop. No hand-rolled orchestration. 3-provider LLM failover: Groq (openai/gpt-oss-120b) → Gemini 2.5 Flash → NVIDIA. Free tiers only. Agent never stops because one provider blinks.

grammy for Telegram. Fastify 5 server. Supabase for memory + reasoning logs. node-cron scheduler: every 6h, Siren reviews open splits and decides — nag, check-in, wait, or nothing. Full reasoning persisted and viewable via /siren why or the per-group dashboard. Custom tools alongside the toolkit: fetch_group_history, get_member_profile, remember_preference, schedule_followup, summarise_debt, telegram_dm — the agent composes them freely. Deterministic guard rails: LLM writes voice, code writes facts. /status and settlement posts are direct DB reads — no hallucinated amounts. Rate limits in handlers, not just prompts: 1 reminder per user per 24h, enforced in telegram_dm before the API call. Dashboard (/siren dashboard) — signed per-group link, 24h expiry, AG Grid tables showing member reliability, splits, and every agent decision with its full step-by-step reasoning.

Challenges

Keeping the agent from inventing amounts or pay links. Fix: strict separation between LLM-written voice and code-written facts, enforced across every surface. Free-tier LLM rate limits during testing. Fix: 3-provider failover chain + 1h prompt cache. End-to-end testing against a live hosted deploy without mocking PayPal. Built a secret-gated test sink in production that captures Telegram DMs for the test harness — 7/7 criteria green against the live Render URL.

Accomplishments

Shipped in 26 days by a solo second-year student. 7/7 success criteria passing end-to-end against live production — not local mocks. Dashboard with full agent-reasoning viewer — judges can inspect every decision the agent made. 269 unit tests, 17/17 PII redaction tests, 9/9 live-URL verification checks.

What’s next

Receipt vision: photo of a bill → line-item split (Gemini Vision + parse_receipt tool). A2A cross-group settlement: two Siren instances negotiating net transfers between groups. Production PayPal (live mode) with KYC-approved merchant accounts.

Built With

Share this project:

Updates

Submission history