Build Log #4: Week 3 — Premium Lumen & Convergence Pillars
Weeks 1 and 2 built the foundation: blog, data pipelines, tools. Week 3 was about differentiation — the premium layer that would separate Picking Solutions from every other market intelligence aggregator.
The Lumen Dashboard
Lumen is the premium intelligence dashboard — the core product proposition. While the daily feed gives you curated headlines, Lumen gives you analysis: digests synthesised by AI, pulse signals with conviction scores, opportunity radar, and a three-column feed combining editions, research summaries, and real-time market signals.
The architecture:
- /lumen — Dashboard overview with key metrics and navigation
- /lumen/feed — Three-column intelligence feed (Today's Picks, Research Digests, Pulse Signals)
- /lumen/stream — The Intelligence Stream: AI-synthesised digests for Official, Research, and Social intelligence
- /lumen/pulse — Signal tracking with thesis, confidence level, and time horizon
- /lumen/opportunities — Market opportunities combining energy and crypto data
- /lumen/reports — Monthly deep-dive reports (AI-generated, human-reviewed)
Each page pulls from different data models but shares a consistent design language: dark card backgrounds, cyan accent colours, metric-dense layouts with clear visual hierarchy.
Intelligence Digests
The centrepiece of Lumen: AI-synthesised digests. The Content Synthesiser (a scheduled task running 3x daily at 06:00, 12:00, and 19:00 UTC) reads approved intelligence items, groups them by type, and generates narrative digests using the RouteLLM API.
Three digest types:
- Official Sources — Reuters, FT, Bloomberg — curated for hard news and data
- Research & Analysis — Carbon Brief, analyst blogs, deep dives — curated for insight
- Social Intelligence — X/Grok-sourced takes on energy, BTC, RWA, AI, DeFi
The prompt engineering was critical. Early digests were too generic — they read like any AI summary. We iterated until the output felt like a briefing: opinionated, cross-referencing sources, highlighting contradictions between analysts, and ending with a "so what" takeaway.
The 48-hour lookback window was a key design choice. Digests don't just cover today — they look back 48 hours to catch items that were ingested overnight or during off-hours. This prevents the morning digest from missing late-night developments.
Pulse Signals
Signals are different from digests. Where digests summarise what happened, signals express a thesis: a directional view with a confidence score (0-100), time horizon, and supporting evidence.
Example: "UK wholesale electricity prices likely to decline 15-20% over next 2 weeks as wind generation forecast exceeds seasonal average. Confidence: 72%. Horizon: 14 days."
Signals are generated during the morning Content Synthesiser run, using the latest digest content plus market data context. They're stored in the PulseSignal table with status tracking (DRAFT → PUBLISHED → ARCHIVED) and optional expiry dates.
The Reports Pipeline
Monthly reports: 5,000-6,000 word deep dives covering each vertical. Generated by the Monthly Report Generator task (1st of each month, 09:00 UTC), then reviewed and published by the admin.
The generation prompt is extensive — it includes the month's digest summaries, signal outcomes, market data trends, and editorial guidelines. The AI produces a structured report with executive summary, vertical-by-vertical analysis, data highlights, and forward outlook.
Reports are generated as drafts. The admin reviews, edits, and publishes. This human-in-the-loop step is deliberate: monthly reports carry the platform's analytical reputation. They need human judgment before going live.
The Convergence Thesis
Why cover energy, finance, and digital assets together? Because the interesting things happen at the intersections:
- Tokenised energy assets — renewable energy certificates on blockchain
- DeFi meets traditional finance — real-world asset (RWA) tokenisation
- AI + energy — data centre power demand, grid balancing automation
- Carbon markets — intersection of energy policy, financial instruments, and crypto settlement
The convergence thesis isn't just branding — it drives data architecture. Sources are tagged by vertical, but the best content often spans two or three. The digest system highlights cross-vertical themes. The pulse signals explicitly look for convergence patterns.
Week 3 Technical Debt
Moving fast meant accumulating debt. By end of week 3:
- Dark mode was broken in multiple places (the prose styling disaster — white text on white backgrounds)
- The admin panel had features nobody used yet
- Image handling was inconsistent — some pages used Next.js Image, others used raw
- TypeScript was set to strict but several
anytypes had crept in
We logged everything in GO_LIVE_BACKLOG.md and kept moving. Shipping beats perfection in a 30-day build.