INSPIRATION
Every morning at approximately 08:00, the President of the United States receives the President's Daily Brief — a classified intelligence document prepared overnight by 400+ analysts, signals officers, and intelligence professionals. It is cross-validated, threat-graded, and delivered with human-level narrative synthesis.
The President makes better decisions because of it.
The rest of the world's 350 million executives, 2 million fund managers, 80,000 board members, and countless policy-makers, researchers, and journalists start their day with a Twitter scroll.
This is not an information scarcity problem. The internet produces over 2.5 million articles daily. It is an intelligence processing problem. No technology has ever translated raw information velocity into structured, graded, decision-ready intelligence — at the scale and price point that individuals and organizations outside of government could access.
We built IBF because four forces finally converged to make it possible:
1. Amazon Nova 2. The Nova 2 model family — with Nova 2 Lite's extended thinking and reasoning, Nova 2 Sonic's speech-to-speech narration, Nova Act's 90%-reliable browser automation, and Nova Multimodal Embeddings' unified cross-modal retrieval — provides, for the first time, a complete intelligence pipeline on a single cloud provider, without stitching together a dozen different vendors.
2. Cost compression. Token costs have fallen to a level where running a full 8-stage pipeline per user costs under $0.20. The economics now support consumer-tier pricing for something that previously required nation-state budgets.
3. News API maturity. GDELT, NewsAPI, Perigon, and a dozen other services now provide programmatic access to 100,000+ news sources in real time — the raw material the pipeline needs.
4. Production-grade orchestration. Amazon Bedrock AgentCore now provides the governance, observability, and deployment infrastructure to run a pipeline like this reliably in production — not just in a prototype.
IBF was born from a simple question: what if the intelligence briefing that makes the most powerful person in the world better at their job was available to everyone, every morning, for the price of a coffee?
WHAT IT DOES
IBF watches the internet on behalf of its users. Overnight. Without being asked. And every morning, it delivers a structured, graded, source-cited intelligence brief — the kind that heads of state, fund managers, and intelligence directors pay hundreds of thousands of dollars a year to receive — to any individual, team, or organization.
It is not a news aggregator. Not a summarizer. Not a smarter RSS reader. It is the only platform that transforms raw information velocity into a decision-ready briefing document through a sequential, multi-stage processing pipeline — and delivers it narrated, interactively, by voice.
The Three Core Experiences
1. The Morning Brief (06:00 Daily) A structured intelligence document covering all active Watch Areas. It opens with a single lead item — the most important development overnight — followed by a Threat Matrix, Watch Area Reports (each with situation + context + signal + source credibility), Emerging Signals (early-stage patterns that don't yet constitute news), cross-domain Analyst Notes where the system identifies connections across separate domains, and a Tracking Update on items flagged in prior briefs.
The user reads it in 7–12 minutes. Or — through Nova 2 Sonic — they hear it during their commute, narrated in a natural human-quality voice, in their language.
2. The Flash Brief (Hourly, On-Demand) A compressed version delivered on demand at any hour or auto-triggered during active hours. Covers only new developments since the previous cycle. Delivered in under 4 minutes from trigger to inbox.
3. The Deep Dive (On-Demand) A long-form intelligence report (1,500–3,000 words) on any single topic, entity, or developing situation. Pulls 72 hours of article volume, historical context from the knowledge store, and forward projections.
Watch Areas
Every account is built around configurable Watch Areas — the intelligence domains IBF monitors on the user's behalf. Users define them in plain English:
- "Monitor Fed rate decisions and their impact on tech valuations"
- "Track EU digital regulation — all legislative developments and enforcement actions"
- "Watch our three main competitors — every product launch, hiring signal, and press mention"
- "Monitor political stability in Southeast Asia for our supply chain exposure"
IBF translates plain-language intent into structured filter configurations, named entity watchlists, and source prioritization — no configuration UI, no boolean query syntax required.
HOW WE BUILT IT — THE RELAY INTELLIGENCE ARCHITECTURE (RIA)
IBF is powered by the Relay Intelligence Architecture — a sequentially orchestrated, fault-tolerant 8-stage pipeline where each Amazon Nova model performs a single, specialized intelligence function and passes its complete structured output as context to the next.
The design mirrors how real intelligence communities work: each layer contributes one expertise. The final stage synthesizes the whole.
WORLD NEWS FEEDS (100K+ sources)
│
▼
┌─────────────────────┐
│ STAGE 1 — SCOUT │ ← Amazon Nova Act
│ Web Harvester │ Autonomous browser workers for
└──────────┬──────────┘ sources without APIs
│ Raw Article Corpus (200–2,000 items)
▼
┌─────────────────────┐
│ STAGE 2 — CLASSIFIER│ ← Nova 2 Lite (Thinking: LOW)
│ Taxonomy Engine │ Domain, entity, sentiment, tone,
└──────────┬──────────┘ geography, event type
│ Tagged Article Objects
▼
┌─────────────────────┐
│ STAGE 3 — RELEVANCE │ ← Nova 2 Lite (Thinking: LOW)
│ Scoring Officer │ Scores against User Context Model
└──────────┬──────────┘ Prunes corpus to top N items
│ Scored + Ranked Articles
▼
┌─────────────────────┐
│ STAGE 4 — VERACITY │ ← Nova 2 Lite (Thinking: LOW)
│ Cross-Validator │ Source credibility, corroboration,
└──────────┬──────────┘ manipulation signal detection
│ Validated Intelligence Items
▼
┌─────────────────────┐
│ STAGE 5 — ASSESSOR │ ← Nova 2 Lite (Thinking: MEDIUM)
│ Risk + Signal │ Threat levels 1–5 + weak signal
└──────────┬──────────┘ detection + temporal urgency
│ Risk-Graded Intelligence
▼
┌─────────────────────┐
│ STAGE 6 — HISTORIAN │ ← Nova 2 Lite + Multimodal Embeddings
│ Context Weaver │ RAG from knowledge store — prior
└──────────┬──────────┘ briefs, events, trend continuity
│ Context-Enriched Intelligence
▼
┌─────────────────────┐
│ STAGE 7 — ANALYST │ ← Nova 2 Lite (Thinking: HIGH)
│ Brief Composer │ Full structured brief synthesis
└──────────┬──────────┘ with cross-domain analytical notes
│ Draft Brief
▼
┌─────────────────────┐
│ STAGE 8 — EDITOR │ ← Nova 2 Lite (Thinking: LOW)
│ Quality Gate │ Citation verification, confidence
└──────────┬──────────┘ scoring, tone compliance
│ Final Brief (Pipeline Confidence Score attached)
▼
DELIVERY ENGINE
┌─────────────────────┐
│ Nova 2 Sonic │ Speech-to-speech narration
│ Email / App / API │ Interactive voice delivery
└─────────────────────┘
HOW EACH AMAZON NOVA MODEL IS USED
Amazon Nova Act → Stage 1 (The Scout)
The best intelligence sources — parliamentary records, niche trade publications, regional newspapers, regulatory portals, central bank websites — do not have APIs. Nova Act solves this with a fleet of autonomous browser workers that navigate real websites the way a researcher would.
Each task is defined in plain English: "Go to the Reserve Bank of India press release page. Find any releases published in the last 6 hours. Extract the title, date, and summary text." Nova Act interprets the live page structure and executes — handling pagination, JavaScript-rendered content, and login-free gated portals — without brittle CSS selectors that break every time a site redesigns.
Nova Act's 90% reliability on browser-based workflows makes this production-viable for the first time. The Scout fleet reaches 80+ sources per Watch Area that no API-only architecture can touch.
Amazon Nova 2 Lite (Extended Thinking) → Stages 2, 3, 4, 5, 7, 8
Nova 2 Lite is the reasoning engine across six of eight pipeline stages. Its 1M-token context window — large enough to hold an entire article corpus plus the User Context Model in a single call — and its configurable Extended Thinking levels (LOW / MEDIUM / HIGH) allow each stage to be tuned precisely for the work it does:
- Stages 2, 3, 4 (Classify, Score, Validate): Thinking at LOW — high throughput, low cost. These stages process hundreds of items and need speed, not depth.
- Stage 5 (Risk Assessor): Thinking at MEDIUM — threat level assignment requires multi-step reasoning across the full validated set to detect cross-article patterns and early warning signals.
- Stage 7 (Brief Composer): Thinking at HIGH — the most cognitively demanding task in the pipeline. The Composer receives the full enriched corpus and must produce a coherent, narratively structured brief with original cross-domain analysis. HIGH thinking is what makes IBF briefs read like something a senior analyst wrote, not a summary generator produced.
- Stage 8 (Quality Gate): Thinking at LOW — a focused audit task: verify every claim traces to a source, flag unverifiable specifics, confirm structural coherence.
Nova 2 Lite's built-in web grounding supplements Stage 6 (Historian) when the knowledge store lacks depth on a specific entity — the stage can anchor itself in current web data without a separate tool call.
Nova Multimodal Embeddings → Stage 6 (Historian) + User Context Model
IBF uses Nova Multimodal Embeddings in two ways that no text-only embedding model could support:
1. The Intelligence Knowledge Store. Every article processed by the pipeline is embedded — including its headline, body text, and any associated images (satellite imagery captions, charts, data visualizations in PDF reports). This creates a single semantic space where text and visual content are retrievable together. When Stage 6 queries for historical context — "what do we know about ECB rate decisions over the last 90 days?" — it retrieves not just text articles but charts and document images processed in prior cycles.
2. The User Context Model (UCM). The UCM — IBF's representation of what matters to a specific user — is maintained as a vector that evolves with engagement. When the user reads a brief item fully, shares it, or provides explicit feedback, those signals are embedded and used to update the UCM. Stage 3 (Relevance Scoring) performs a semantic similarity operation between each article embedding and the UCM vector as part of every scoring cycle.
Amazon Nova 2 Sonic → Voice Brief Delivery
This is IBF's most visible consumer differentiator and, in testing, the feature that changed how people described the product.
Every morning brief can be delivered as a narrated voice session. The user opens IBF on their phone at 06:00 and hears their brief — not a text-to-speech robot reading words on a screen, but Nova 2 Sonic understanding the material and delivering it with natural cadence, emphasis, and turn-taking. The experience is closer to a private briefer than an app.
Nova 2 Sonic's polyglot voice capability means a user in Germany or Brazil hears their brief in their native language, with the voice switching naturally when quoting a foreign official mid-sentence.
Nova 2 Sonic's asynchronous tool processing enables a live interactive mode: the user can interrupt mid-brief and ask "Tell me more about item three" or "What does this mean for the bond markets?" — and Nova 2 Sonic handles the background retrieval and resumes narration simultaneously, without pause, without losing context.
The 35-minute morning commute becomes an intelligence briefing session.
AWS ARCHITECTURE
| Component | AWS Service | Role |
|---|---|---|
| All Nova Models | Amazon Bedrock | Model inference — all 8 pipeline stages |
| Pipeline Deployment | Amazon Bedrock AgentCore | Production fleet governance + observability |
| Browser Automation | Amazon Nova Act + AgentCore Runtime | Stage 1 browser worker fleet |
| Voice Sessions | Nova 2 Sonic via Bedrock Streaming API | Real-time voice brief delivery |
| Knowledge Store | Amazon OpenSearch Serverless | Intelligence history + UCM embeddings |
| Brief Scheduling | Amazon EventBridge | Morning brief triggers per user timezone |
| Stage Workers | AWS Lambda | Serverless execution for Stages 2–8 |
| Corpus Hand-off | Amazon SQS | Article objects passed between stages |
| User Profiles | Amazon DynamoDB | Subscriptions, UCM state, brief history |
| Archive | Amazon S3 | Historical brief storage + PDF generation |
| Email Delivery | Amazon SES | Morning brief + Flash brief email delivery |
| Push Alerts | Amazon SNS | Flash brief mobile notifications |
| API Layer | Amazon API Gateway | REST interface for web app + integrations |
| CDN | Amazon CloudFront | Global latency reduction |
| Authentication | Amazon Cognito | User auth + session management |
| Monitoring | Amazon CloudWatch | Pipeline latency + stage performance |
THE RELAY MECHANISM — WHY SEQUENTIAL MATTERS
The stages do not run in parallel. They run in relay — each stage's complete structured output becomes the total input for the next. This is a deliberate architecture decision, not a constraint.
Why relay over parallel?
Stage 7 (Brief Composer) doesn't just see a relevance score. It sees the original article, every classification tag, the exact relevance rationale, the veracity score, the risk level, and the historian's context block. That accumulated depth is what produces analysis rather than summaries.
Each stage's system prompt is entirely focused on its one function. Specialization produces better outputs than a single generalist prompt trying to do everything.
At every relay point, the structured output is logged. Users can click into any brief item and trace the complete pipeline: which source → how it was classified → why it scored relevant → what risk level was assigned → what historical context was added → how it was synthesized → what the quality gate flagged. This audit trail is what enterprise risk teams and compliance officers require before trusting a system like this.
CHALLENGES WE RAN INTO
1. The 55-Minute Problem
The morning brief runs overnight — 6–8 hours of processing time available. The hourly Flash Brief has 55 minutes and must complete in under 4. Running 8 full-depth stages sequentially in that window was not viable with naive implementation.
The solution: a tiered execution model. For hourly briefs, Stages 2–5 operate against a pre-warmed context cache — article embeddings and UCM vectors computed in the morning cycle are already stored. Only the delta (articles published since the last cycle) runs through the full pipeline. Stages 7 and 8 produce a Flash Brief format — compressed, high-threat-level items only. Result: 3 minutes 47 seconds average pipeline time in testing.
2. Hallucination in Stage 7 (Brief Composer)
Running at HIGH extended thinking, the Brief Composer is the most capable but also most prone to generating plausible-sounding conclusions that aren't directly traceable to any source article.
The solution: Stage 8 (Quality Gate) performs a structured citation audit. Every factual claim in the draft brief is extracted as a typed list and semantically matched against the article corpus. Claims without a match above a confidence threshold are either removed or flagged with a "UNVERIFIED SIGNAL" marker rather than presented as established fact. Nova 2 Lite's 1M-token context window makes this viable — the entire article corpus and the draft brief fit in a single context call, enabling full-document citation grounding without chunking.
3. Training the Scout Browser Fleet
Browser task definitions that worked on one outlet's DOM failed on another. Brittle source-specific scripts were unsustainable at scale.
The solution: a task abstraction layer — generic templates (extract headlines from publication, monitor document release page, search with keywords) that Nova Act interprets contextually per source. The template defines intent; Nova Act reads the live page and determines execution. This removed per-source maintenance entirely.
4. Context Accumulation Across Stages
As the relay progresses, accumulated context grows. By Stage 7, the input includes the full article corpus plus six prior stages' outputs. Without careful JSON schema design, context bloat made later stages slow and expensive.
The solution: each stage outputs a strictly typed, compressed JSON envelope — minimum fields needed for the next stage, with full content archived to S3 and referenced by key rather than included inline. Stage 7 retrieves full article text from S3 only for items it needs to cite.
5. Voice Position State
When a user interrupts the voice brief mid-sentence and says "go back to item two," Nova 2 Sonic needs to know where item two is in the brief structure. This required maintaining a brief position state object alongside the Sonic session, updated in real time as the narration progresses.
ACCOMPLISHMENTS WE'RE PROUD OF
1. A Complete 8-Stage Pipeline — End to End
A full pipeline from news ingestion to final brief delivery — using four distinct Amazon Nova products — shipped as a working system in one hackathon cycle. Not a prototype of one stage, not a demo of the voice layer in isolation. The complete relay, running.
2. Voice Intelligence Briefing via Nova 2 Sonic
No existing news or intelligence platform delivers a narrated, interactive briefing. Not Bloomberg, not Feedly, not any of the dozens of news aggregators that came before. The combination of structured brief format + Nova 2 Sonic's natural speech + interruption handling creates a product experience that is qualitatively different from reading. When the first voice brief played back in testing, the team's reaction was unanimous: this is a different product.
3. The Pipeline Confidence Score (PCS)
Every IBF brief includes a PCS — a composite quality metric derived from Stage 8's audit: source credibility average, corroboration count per item, hallucination flag rate, editorial coherence score. The number is visible to the user. It creates a transparent trust signal that builds credibility over time. Fund managers and policy teams in our beta group named this as the feature that made them trust the output enough to act on it.
4. Sub-4-Minute Hourly Brief
A 3:47 average pipeline time for the hourly Flash Brief means IBF serves time-critical decisions, not just morning routines. A fund manager can trigger a brief the moment a market-moving event breaks and receive structured, source-cited intelligence in under 4 minutes.
5. Cross-Modal Intelligence Retrieval
Using Nova Multimodal Embeddings to unify text articles, document images (charts, tables, infographics from regulatory filings), and satellite imagery captions in a single retrieval space was an architectural choice that significantly improved Stage 6's historical context — particularly for financial and geopolitical intelligence where a chart is often more information-dense than the article describing it.
6. Amazon-Native Full Stack
Every layer — browser automation, reasoning, voice delivery, vector retrieval, scheduling, delivery, monitoring — runs on AWS. No hybrid infrastructure, no third-party LLM vendors, no cross-cloud dependencies. This is not a demonstration of what Amazon Nova can do. It is a production-ready system built on what Amazon Nova does.
WHAT WE LEARNED
1. Thinking Budget Is a First-Class Architecture Decision
Calibrating extended thinking level per stage — LOW for throughput tasks, HIGH for synthesis — contributed more to output quality and cost efficiency than any model selection choice. This belongs in the architecture diagram, not as a tuning detail.
2. Nova Act Changes the Question About Data Access
Building the Scout fleet required abandoning the assumption that data access is an API problem. Nova Act changes the question from "does this source have an API?" to "can a person navigate this website?" If yes, Nova Act can handle it. This unlocks access to a class of intelligence sources — government portals, parliamentary records, regulatory databases, regional outlets — that have never been reachable in real-time intelligence pipelines before.
3. Voice Delivery Changes the Product Category
Reading a brief and hearing a brief activate different cognitive modes. For time-constrained executives, voice delivery is not a feature — it is the primary value proposition. Nova 2 Sonic's quality was the difference between this feeling like a genuine product and feeling like a gimmick. Earlier text-to-speech systems would not have cleared that bar.
4. Sequential Relay Outperforms Parallel Merging for Synthesis Quality
Early prototypes ran stages in parallel and merged outputs. Output quality was measurably and consistently lower. The relay architecture — where each stage receives full prior stage context — produced better synthesis because the Brief Composer understood why each item was scored, validated, and risk-graded before writing about it. Context depth produces analysis. Context fragmentation produces summaries.
5. Structured Schema Is the Pipeline's Connective Tissue
Requiring strict JSON schema compliance at every relay point — validated by the orchestration layer before passing to the next stage — eliminated an entire class of mid-pipeline failures. Stages don't fail silently; schema violations surface immediately and are recoverable. This principle scales to any multi-stage processing pipeline.
WHAT'S NEXT FOR IBF
Q2 2026 — Enterprise Tier Dedicated enterprise accounts with unlimited Watch Areas, custom proprietary source lists, the Briefing Room (collaborative annotation, escalation, and action assignment within teams), and board-ready weekly Intelligence Digest reports in designed PDF/PowerPoint format. Target customers: Fortune 500 strategy teams, investment funds, policy advisory firms.
Q2 2026 — Amazon Q Integration For enterprise users, IBF connects to Amazon Q to bridge external intelligence (what's happening in the world) with internal knowledge (what the organization knows, has decided, and is planning). A brief that surfaces relevant internal documents, prior decisions, and team context alongside world events becomes dramatically more actionable than either source alone.
Q3 2026 — Nova Forge Fine-Tuning Using Nova Forge to fine-tune Stage 7 (Brief Composer) on a curated corpus of high-quality intelligence analysis — published think-tank reports, historical Presidential Daily Brief summaries, financial analyst notes — to produce analytical prose calibrated specifically to intelligence briefing conventions. This custom fine-tuned composer becomes a proprietary model asset. It is the moat that no competitor can replicate from outside the AWS ecosystem.
Q3 2026 — Amazon Connect Voice Delivery Using Nova 2 Sonic's native Amazon Connect integration, IBF will offer a phone-call delivery option for enterprise users: every morning, IBF calls the executive and delivers their brief interactively. The user can ask follow-up questions, drill into items, and create action items — all by voice, with no app required. This reaches the class of senior executive who will never open another application.
Q4 2026 — Developer API An open API allowing any third-party platform — media companies, banks, telcos, productivity tools — to request a IBF brief via REST. White-label intelligence infrastructure. The brief becomes embeddable.
2027 — Global Language Expansion Nova 2 Sonic's seven-language polyglot voice capability becomes the foundation for IBF's expansion into non-English markets. Briefs generated, narrated, and delivered in Spanish, French, Portuguese, German, and Hindi — with source monitoring in corresponding languages via GDELT and regional APIs. Target markets: Latin America, Western Europe, India, Southeast Asia.
Log in or sign up for Devpost to join the conversation.