# GM FM
### 1. What We’re Building
We’re launching the first always-on, talk-only crypto radio stream. Every news bulletin, deep-dive segment, and ad read is generated in real time by a coordinated system of AI agents — with <2 minutes from signal to stream.
#### Core Features:
* Real-time response: Breaking news becomes audio within 2 minutes
* Synthetic voices only: No humans on-air
* 24/7 programming: Short bulletins, longer analytical shows, adaptive content
* No accounts, no tracking: Anonymous by design
* On-chain ad billing: Sponsors stay pseudonymous, all payments via Base L2
***
### 2. How It Works — The AI Agent Stack
Each core function is handled by an isolated AI-powered service. Together, they maintain the flow of live crypto coverage, entirely autonomously.
#### 🧠 News Agent
* Scrapes on-chain alerts, RSS feeds, X/Twitter firehose
* Filters sources and deduplicates events
* Converts stories into radio copy via GPT-4o
* Publishes news.item.created to the pipeline
* Latency target: first radio script within 30 seconds of signal
#### 🎬 Showrunner
* Manages the minute-by-minute program clock
* Chooses show format: short (90s), medium (15m), or long (60m)
* Picks synthetic hosts, places ads, and issues segment requests
* Writes schedule changes to programme\_history for auditability
* Reacts in real time to sentiment signals and breaking events
#### 🗣 Voice Synth Service
* Sends each script line to ElevenLabs, receives MP3 in <3s
* Stores audio in S3 and publishes audio.line.ready
* Falls back to cached lines if synthesis fails
* Streams content via HLS in 6-second chunks
#### 💬 Sentiment Agent
* Monitors live listener chat
* Scores messages for emotion, toxicity, and relevance
* Updates real-time popularity metrics per host and topic
* Feeds the Showrunner with swap signals for unpopular content
* Embedding rate capped at 100 msg/s
#### 💰 RTB Ad Engine
* Real-time bidding via WebSocket; CPM-based auctions
* On-chain escrow contract in USDC on Base L2
* GPT-4o writes 30s native ad reads, Voice Synth voices them
* On playback, a signed impression receipt is posted on-chain
* All ad reads audibly marked as “Sponsored”
#### ✅ Moderation & Compliance
* Runs all chat and script content through OpenAI Moderations + AWS Comprehend
* Blocks or delays toxic content
* Injects “Sponsored” tags automatically
* Logs every paid placement on-chain for full transparency
* Borderline content is quarantined in an SQS Dead Letter Queue
***
### 3. The Full Production Pipeline
| Stage | Time Budget |
| ---------------------------- | ------------- |
| News detection → script | ≤ 30 seconds |
| Segment creation & host pick | ≤ 500 ms |
| Voice generation per line | ≤ 3 seconds |
| First audio chunk available | ≤ 120 seconds |
| Long-form shows (full) | ≤ 4 minutes |
HLS streaming updates every 6 seconds with fresh audio segments.\
***
### 4. Sponsorship Workflow
#### 4.1 Campaign Setup
* Sponsor creates campaign via web portal: set CPM, daily cap, and prompt hints
* Portal returns a unique Base L2 deposit address
* USDC received → campaign funded → eligible to bid in real time
#### 4.2 Real-Time Bidding
* Each ad slot includes topic + CPM floor
* Sponsors submit bids via WebSocket (CPM + max impressions)
* Redis ZSET selects highest bidder; second-price logic optional
* GPT-4o writes native ad read → voiced and stitched into stream
#### 4.3 Proof & Escrow
* Each playback creates a SHA-256 hash (segment ID + timestamp)
* Hash is signed and posted on-chain via recordImpression()
* Contract debits escrow; emits ImpressionReceipt
* Gas too high? Queue for batch posting later (90% gas savings)
#### 4.4 Budget Monitoring
* Alerts when campaign drops below 10% remaining
* Auto-refund on campaign end if unused funds remain
* Withdrawals require 2-of-3 multisig (system key + ops hot-key)
#### 4.5 Reporting
* CPM is USD-pegged, converted via Chainlink
* GraphQL subgraph indexes all receipts
* Real-time dashboards + CSV export (with tx\_hash for self-audit)
***
### 5. Infrastructure Overview
Everything runs on AWS with automated deployment and scaling.
#### Compute & Orchestration
* Amazon EKS for container orchestration
* Spot instances for all agents except RTB Engine
* Karpenter + Cluster Autoscaler handle node scaling
#### Networking
* Istio service mesh: mTLS, retries, traffic mirroring
* API Gateway for WebSocket endpoints (chat + RTB)
* Layer-7 routing inside mesh
#### Storage & Messaging
* S3: MP3 storage, HLS segments, logs
* Redis: sentiment scores, RTB bids, deduplication filters
* Postgres: program clock, ad ledger
* OpenSearch: news embeddings + similarity search
* EventBridge: core pub/sub event bus
* SQS: queues for moderation, synthesis, encoding
#### Media Pipeline
* FFmpeg jobs generate audio chunks and write directly to S3
* CloudFront delivers HLS with adaptive caching (12h for .ts, 1m for playlists)
#### Observability
* Prometheus + Grafana: latency, token usage, queue depth
* AWS CloudWatch for logs + alerts
* AWS X-Ray: trace full pipeline from news trigger to audio stream
***
### 6. Performance Targets
| Metric | Goal |
| ------------------------------- | ------------------------------ |
| Headline → air (short) | ≤ 120s (P95) |
| Full show (long-form, 15–60min) | ≤ 4 min |
| Ad bid → playback | ≤ 3 seconds |
| Listener concurrency | 100,000+ |
| Stream availability | 99.9% monthly |
| Recovery time objective (RTO) | 30 minutes |
| Recovery point objective (RPO) | 1 hour (Postgres + S3 backups) |
***
### 7. Principles & Design Philosophy
* Real-time media: Crypto moves fast — the stream should too.
* Modular agents: Each component is containerized and hot-swappable.
* Anonymity-first: No cookies, no accounts, no personal data collected.
* On-chain accountability: Every ad logged immutably.
* Resilient by default: Multi-AZ, autoscaled, backed up.
***
For technical specs, developer docs, and deployment guides, see:
* Agent Architecture
* Ad System & Smart Contracts
* Deployment & Observability
* Stream Format & HLS Specs
End of section.