Gáldr × Viska — AI Intelligence Department Roadmap

Document for: CIO

Purpose: Goals, methodology, infrastructure decisions, and timeline for the 6-month engagement.

The Goal

Build Viska a fully operational AI intelligence department over six months. At the end, Viska owns every component — code, data, infrastructure — and can operate it independently or with ongoing support.

Our Methodology

How We Build AI Systems

The AI industry moves fast. New models ship quarterly. New tools appear weekly. The temptation is to grab the latest thing and wire it up. This produces systems that break every time the landscape shifts.

We work differently. Our methodology is built on three phases that repeat at every level of the system — from the overall architecture down to individual tool selection.

Phase 1: Research Before Building

Before writing a line of code, we research the landscape systematically. For the Trading War Room alone, we conducted 9 formal research investigations over two weeks, covering:

• 11 external services audited for integration potential
• 2 TradingView MCP projects evaluated tool-by-tool (one declined, one selectively adopted)
• MCP vs CLI architecture compared across 25 test runs for token cost and reliability
• External API integration patterns mapped (Grok, FRED, ECB, Ollama)
• Security evaluation of supply chain dependencies
• Knowledge retrieval patterns benchmarked (index vs RAG)

Every finding is documented. Every decision has a rationale. Nothing is adopted without evaluation. Nothing is discarded without a written reason.

Phase 2: Evaluate Against Criteria

Every tool, model, or integration is measured against six criteria before it enters the system:

Case study — TradingView integration:

The CIO's actionable plans reference TradingView for charting and technical analysis. We evaluated two TradingView MCP projects in detail:

Project 1 (browser automation approach) — Declined. Five structural kill factors:

Project 2 (Python library approach) — Selective adoption for later phase. No browser automation. Backtesting, screening, and multi-timeframe analysis tools fill a genuine gap. But: sentiment tools use keyword counting (not real NLP), news tools duplicate our existing pipeline, and the underlying TradingView library is 4 years stale. We adopt only the tools that add value, skip the rest.

What we built instead: A three-layer architecture using TradingView's own official features:

• TradingView Webhooks for signal delivery (official, supported)
• TradingView Lightweight Charts for dashboard display (MIT-licensed, ToS-compliant)
• FRED MCP + QuantOracle for the quantitative macro layer (63 deterministic tools + 800,000+ macro data series)

This is the methodology in action. We don't adopt or reject tools based on popularity. We evaluate systematically, document the rationale, and build the architecture that serves the actual need.

Phase 3: Orchestrate, Don't Accumulate

Recent research from Stanford and Tsinghua (March 2026) quantified what we practice: the same AI model delivers a 6x performance difference depending on how it's orchestrated. The model is a commodity. The orchestration layer — how agents coordinate, remember, enforce boundaries, and adapt to the domain — is the asset.

This field is called harness engineering. Three principles:

Subtract before you add. The instinct is always more tools, more agents, more complexity. The evidence shows the opposite. One company removed 80% of its tools and got better results. We build focused systems with clear boundaries. Each agent has a defined domain — what it can access, what it can modify, what it must escalate.

Build the layer that survives. Models change quarterly. Prompting techniques that work today break tomorrow. But the orchestration layer — the skill system, the memory architecture, the governance protocols — transfers across models. When a new generation ships, the system gets stronger without rebuilding. Your investment compounds instead of depreciating.

Govern first, optimize second. Speed without governance produces systems nobody trusts. Credential isolation, audit trails, scope boundaries, and human approval gates are structural — built in from day one, not added after an incident. This is what makes AI systems safe for institutional use.

The Research-to-Decision Pipeline

Every component in the system follows the same path:

Research → Evaluate → Decide (adopt / selective adopt / defer / decline) → Document → Build

The documentation is not bureaucracy — it's institutional memory. When a question comes up six months later ("why don't we use X?"), the answer exists with the full reasoning chain. Decisions don't get relitigated without new information.

Six Goals — Six Months

Goal 1: Macro Research Intelligence

Months 1–2

Outcome: The team queries the entire global macro landscape — broker research, central bank minutes, earnings transcripts, policy documents — from a single interface. Cross-asset macro narratives synthesized daily.

What changes:

• Daily brief arrives synthesized before the team sits down — themed, ranked, high-impact findings flagged
• Any team member queries the full corpus and gets sourced, cited answers in seconds
• Real-time narrative tracking surfaces sentiment shifts on tokens, macro themes, and geopolitical triggers as they happen
• Corpus expands beyond broker PDFs to primary sources: Fed, ECB, BoJ minutes, earnings transcripts, economic data
• Accessible from any workspace — Slack, Claude Desktop, or any MCP-compatible tool

Goal 2: Multi-Agent Market Surveillance

Months 2–3

Outcome: Specialized AI agents run 24/7, each watching a different domain. They monitor, analyze, and report proactively.

Agents deployed:

• Crypto surveillance — on-chain metrics, exchange flows, funding rates, anomaly detection
• Macro sentiment — cross-asset radar, conviction scoring across key themes (AI capex, commodity supercycle, deglobalization)
• Geopolitical risk — regulatory filings, sanctions, elections, policy shifts

What changes:

• 24/7 coverage without anyone staying up
• Consensus divergence flagged — when data contradicts broker opinion, the team knows immediately
• Every finding logged to institutional memory for later retrieval

Goal 3: Trading War Room

Months 2–4

Outcome: A four-layer trading environment with live data, technical analysis, trade journaling, and mechanical risk enforcement. Every order requires human approval. Every trade has a written thesis.

Architecture:

Order flow — defense in depth:

Analyst thesis → Risk qualification → Operator approval → Execution submission → Alpaca

Every step produces a structured record. No step can be skipped.

Risk framework:

Non-negotiable:

• Human approval on every order — schema-enforced at code level
• 90-day paper trading before any live capital discussion — hard gate
• Phase 1: long-only, no options, no leverage, no crypto, no autonomous execution

Goal 4: Investor Communications & Content

Months 4–5

Outcome: Monthly LP letters, performance reports, and thought leadership content draft automatically from actual research and positions. The team reviews and approves — never starts from a blank page.

Components:

• LP reporting — monthly letters from portfolio data and market context, performance attribution auto-generated
• Thought leadership — 2-3 pieces/week in Icelandic and English, targeting Nordic and European family offices. Viska's macro thinking reaches the audience that should hear it — consistently
• Meeting intelligence — before family office meetings: deep prospect research (portfolio, investments, board seats, philanthropic interests). After: personalized follow-up mapping Viska's offerings to portfolio gaps

Goal 5: Scenario Modeling & Signal Generation

Month 5

Outcome: Overnight scenario analysis and proprietary signal generation from Viska's own macro framework.

Components:

• Scenario lab — deep analysis runs overnight. "What happens to our crypto allocations in stagflation with a weakening dollar?" Structured output by morning: probability assessment, portfolio impact, recommended actions, evidence
• Signal generator — open-source models (Gemma, DeepSeek) fine-tuned on Viska's IC minutes, research output, and historical positions. Generates quantitative signals reflecting the team's qualitative thesis. Runs on Viska infrastructure — IP fully protected

What changes:

• "Leading macro thinking" becomes provable and presentable to LPs
• Qualitative conviction becomes backtestable alpha
• The team's analytical edge compounds — it's a trained model, not tribal knowledge

Goal 6: Distribution Intelligence & Handoff

Month 6

Outcome: Continuous partner surveillance and full system handoff.

Components:

• Partner intelligence — weekly scans of targets across Luxembourg, Nordics, Europe: regulatory filings, AUM shifts, personnel changes, news. Due diligence becomes continuous, not episodic.
• System handoff — complete codebase on Viska GitHub, architecture documentation, runbooks, health dashboard, knowledge transfer

Ownership at handoff:

• All code on Viska GitHub
• Knowledge base on Viska Supabase
• Claude subscription on mimir@viskasjodir.is
• Dashboard on Viska domain
• Nothing locked in — any developer can extend the system
• Optional maintenance retainer, not required

The AI Department — Agents & Capabilities

This is the complete inventory of agents, automation pipelines, and capabilities being built. The CIO can audit each profile against the goals above.

Trading War Room Agents

Four agents operate in coordinated sessions. Each has a defined role, specific tools, and strict scope boundaries.

Scheduled sessions:

• Pre-market review: 08:30 ET daily
• Midday check: 12:30 ET daily
• End-of-session synthesis: 16:15 ET daily
• Alert-driven sessions: triggered by L2/L3 alerts at any time

Later-phase agents (Phase 1D):

Surveillance Agents (24/7)

Intelligence & Communications Agents

Automation Pipelines (n8n)

Seven workflows running on self-hosted n8n, automating data movement and processing:

Capability Matrix

What the team can do at each milestone:

Tool Stack

What We Use

Every tool was evaluated against the six criteria described in Our Methodology.

Agent Tools (MCP Servers)

Infrastructure

Analysis Libraries

What We Evaluated and Declined

Research That Shaped the Architecture

9 formal investigations produced key architectural decisions:

Security Architecture

• Credential isolation — API keys never enter the AI reasoning layer. Injected at runtime through a sandboxed process. The AI uses credentials. It cannot see them, store them, or leak them.
• Scope boundaries — each agent has a defined domain, enforced mechanically. Research agents cannot touch financial data. Communication agents cannot execute trades.
• Audit trails — every decision, action, and data access is logged and traceable.
• Human authority — no consequential action without human approval. Permanent design principle.
• Browser-side security — credentials never reach the dashboard. The FastAPI proxy handles authentication server-side. Dashboard is read-only.

Timeline