Gáldr × Viska sjóðir — Building an AI Intelligence Department

Companion document to the 6-Month Roadmap. Prepared for Guðlaugur Steinarr Gíslason (CIO).

Why This Document

The roadmap describes what we propose to build. This document explains why we build it the way we do — the reasoning behind architectural decisions, the landscape of available tools, the security model that protects the fund, and why this is careful, phased work rather than a quick installation.

The Landscape: What Exists Today

AI-assisted trading and investment research is no longer experimental. As of April 2026, three categories of tools are available to any fund manager:

1. Cloud Automation (Anthropic Claude Routines)

Anthropic released Claude Routines on April 14, 2026. A routine is a saved Claude session — a prompt, connected repositories, and MCP connectors — that runs autonomously on Anthropic's cloud infrastructure. Routines can trigger on a schedule (hourly, daily, weekly), via API call (HTTP POST from any system), or in response to GitHub events (pull requests, releases).

This is a real product with real capabilities. A fund could set up a Claude Routine today that:

• Runs every morning and summarizes overnight market news
• Triggers when a TradingView alert fires and logs the signal
• Scans merged code changes weekly and flags documentation drift

What Routines cannot do: coordinate multiple agents with different roles. A Routine is a single session — one agent, one prompt, one execution. It has no persistent memory across runs (the repo is cloned fresh each time). It has no awareness of what other Routines are doing. There is no risk enforcement across sessions, no portfolio-level awareness, no team dashboard.

A Routine is a capable individual contributor. What we are building is a department.

2. Open-Source Skill Libraries

The open-source community has produced substantial trading skill libraries for Claude. The most mature is a repository of 40+ pre-built investment skills covering:

These skills are well-engineered. Most require only a free-tier financial data API (250 requests/day). Some include automated self-improvement — a daily process that scores each skill and auto-generates pull requests to improve those below threshold.

What skill libraries cannot do: reason about your fund. They are generic — designed for individual retail investors running one analysis at a time. They have no concept of Viska Crypto vs Viska Macro, no awareness of your portfolio positions, no connection to your broker research corpus, no compliance awareness for sub-threshold AIFM limits. They produce analysis for "a stock." We produce intelligence for your fund.

3. Autonomous Trading Agents

Fully autonomous trading frameworks exist. One major open-source platform supports stock, crypto, and derivatives execution across multiple brokers with a three-layer pipeline: signal generation, LLM-based decision making, and automated order execution.

We do not build this. Here is why:

• Performance reality: 92.4% of automated traders on prediction markets lost money. Claimed returns of 59% annualized reflect cherry-picked backtest windows. Realistic benchmarks from peer-reviewed research show 10-20% excess returns on favorable markets — and this before accounting for the judgment errors that LLM interpretation of natural-language instructions introduces.

• Security catastrophe: In February 2026, the largest supply-chain attack on an AI agent ecosystem hit this category. 2,419 suspicious community-contributed skills were removed from the marketplace. Of those, 1,184 were actively distributing wallet-stealing malware. One malicious package had 14,285 downloads before detection. Separately, 21,000 publicly accessible agent instances were found completely unauthenticated — API keys, wallet access, and chat logs exposed to the internet.

This is not a hypothetical risk. This is what happens when trading infrastructure is assembled from unvetted components without security architecture.

What We Build Instead

We build a coordinated intelligence department where every component has a defined role, enforced boundaries, and auditable behavior. The distinction from stock tools:

Single Agent vs Coordinated Team

A Day in the System

Before the team sits down in the morning:

1. 06:00 ET — Overnight synthesis pipeline processes central bank publications, broker research uploaded to the corpus, and macro data from 800,000+ FRED series. Results stage to the portal.
2. 07:00 ET — Daily intelligence brief generates. Themed report covering overnight developments, high-impact flags, sentiment shifts. Delivered to Slack and the portal.
3. 08:00 ET — Watchdog pipeline activates. Position monitoring begins at 5-minute intervals. Alert hierarchy: L1 (informational), L2 (action suggested), L3 (immediate attention).
4. 08:30 ET — Pre-market trading session fires. Four agents convene: Operator reviews overnight intelligence, Analyst presents trade ideas with quantitative confirmation, Risk Manager evaluates against the 8-item checklist, Execution Agent validates order parameters.
5. Throughout the day — TradingView signals arrive via webhook, are validated, enriched with FRED macro context and QuantOracle metrics, logged to the database, and surfaced to the portal with notification.
6. 12:30 ET — Midday session. Position review, thesis update, new signals evaluation.
7. 16:15 ET — End-of-session. Daily performance attribution. Trade journal entries. Risk framework reconciliation.
8. Friday — Weekly report: QuantStats performance tearsheet, P&L attribution, narrative summary of the week's trading decisions and outcomes.

Nothing in this cadence is possible with a single Claude Routine or a library of generic skills. It requires agents that know their roles, remember their context, and operate within enforced boundaries.

What We Adopt from the Ecosystem

We are not building everything from scratch. The ecosystem has strong components — we evaluate them rigorously and adopt what passes our six-criteria framework:

What we do NOT adopt: browser automation tools (ToS violation, fragile), unvetted community skill packages (supply chain risk), autonomous execution frameworks (judgment error risk), full tool suites without filtering (token waste — GitHub MCP burns ~55,000 tokens loading 43 tools).

What We Build Bespoke

Security as Architecture

Security is not a feature we add at the end. It is the foundation on which everything else is built.

The Threat Model for AI-Augmented Funds

Credential Isolation in Practice

This is worth explaining because it is the question most CIOs ask: "How can AI trade on our behalf without seeing our API keys?"

┌─────────────────────────────────────────────┐
│  AI Reasoning Layer                         │
│  ┌─────────┐  ┌─────────┐  ┌────────────┐  │
│  │ Analyst  │  │  Risk   │  │  Operator  │  │
│  │         │  │ Manager │  │            │  │
│  └────┬────┘  └────┬────┘  └─────┬──────┘  │
│       │            │             │          │
│  "Place buy order for AAPL, 50 shares"      │
│       │            │             │          │
├───────┼────────────┼─────────────┼──────────┤
│  Runtime Sandbox (credential injection)      │
│       │                                      │
│  ┌────▼──────────────────────────────────┐  │
│  │  Execution Agent                      │  │
│  │  Validates: operator_approved ✓       │  │
│  │            buy_only ✓                 │  │
│  │            market_hours ✓             │  │
│  │            size_validated ✓           │  │
│  │                                       │  │
│  │  → Calls Alpaca API                   │  │
│  │    (key injected at runtime,          │  │
│  │     never visible to the agent)       │  │
│  └───────────────────────────────────────┘  │
└─────────────────────────────────────────────┘

The agent says "place this order." The sandbox injects the API key into the outbound request. The agent never sees the key value. If the agent's conversation were leaked, exfiltrated, or logged — there are no credentials in it.

Why Six Months

Each phase builds on the previous. Skipping phases creates an intelligence system that has no foundation to stand on.

Month 1: Foundation — Macro Research Intelligence

What we build: The research corpus. Every subsequent agent depends on this — the trading analyst needs macro context, the risk manager needs economic data, the LP communications need sourced intelligence.

Why it takes a month:

• Broker research corpus must be ingested, indexed, and made queryable. This is 400,000+ words of existing research that needs to be structured so agents can retrieve specific findings, not just keyword-match.
• Central bank pipelines (Fed FOMC, ECB rates) need to be built and validated against the official sources.
• FRED integration (800,000+ macro series) needs to be configured for the specific series relevant to Viska's holdings.
• MCP workspace access needs to be established so any team member can query the intelligence layer from their own environment.
• Daily intelligence brief needs to be tuned — the first iterations will be too broad or too narrow. Calibration requires iteration with the team.

What the team gets at end of Month 1: A queryable research corpus. A daily intelligence brief delivered every morning before the team sits down. Central bank data processing on autopilot. Macro data accessible from any workspace.

What cannot happen yet: Trading, surveillance, automated communications. These all depend on the research foundation being solid.

Month 2: Awareness — Multi-Agent Surveillance

What we build: Three autonomous surveillance agents running 24/7 — crypto, macro sentiment, geopolitical risk. Alert pipeline.

Why it depends on Month 1:

• Surveillance agents need the research corpus as context. A crypto surveillance agent without macro awareness produces alerts without context ("BTC down 5%" is noise; "BTC down 5% ahead of FOMC meeting with hawkish expectations" is intelligence).
• Sentiment analysis replaces keyword counting with NLP-based conviction scoring. This requires the Grok integration (Responses API, not the deprecated Chat Completions) and tuning against known market narratives.
• The alert hierarchy (L1 informational, L2 action suggested, L3 immediate attention) needs calibration based on Month 1's baseline of what "normal" looks like.

What the team gets at end of Month 2: Three agents watching markets around the clock. Anomaly detection. Sentiment scoring. Geopolitical risk monitoring. Alert pipeline delivering prioritized notifications.

What cannot happen yet: Trading. Paper trading requires the surveillance layer to be stable — the trading agents use surveillance output as input.

Month 3: Operations — Trading War Room

What we build: Four-agent trading operation. Paper trading on Alpaca. Risk framework enforcing. 7-page dashboard.

Why it depends on Months 1-2:

• The Trading Operator synthesizes surveillance output (Month 2) and research intelligence (Month 1) to form a thesis. Without these inputs, the operator has nothing to synthesize.
• The Technical Analyst uses FRED macro data (Month 1) and QuantOracle metrics for quantitative confirmation. These tools are wired in Month 1.
• The Risk Manager's 8-item checklist references portfolio context that the surveillance layer tracks.
• The dashboard displays data that the automation pipelines (Month 1-2) produce. Building the dashboard first would show empty pages.

The 90-day paper trading gate:

Paper trading begins in Month 3 and runs for 90 days — through Month 5. This is a hard gate: no live capital until the paper trading record demonstrates:

During paper trading, the system operates exactly as it would with live capital. The team reviews trades, evaluates thesis quality, calibrates risk parameters, and identifies failure modes — all without financial exposure.

What paper trading teaches:

• Which risk parameters are too tight (blocking good trades) or too loose (allowing bad ones)
• Whether the alert hierarchy correctly prioritizes events
• How the team actually uses the portal and Slack integration
• Which surveillance agents produce actionable intelligence vs noise
• What the daily operational cadence should look like

This learning is impossible to simulate. It requires real market conditions, real decisions, and real time.

What the team gets at end of Month 3: An operational trading system in paper mode. A 7-page dashboard showing portfolio, positions, strategy, analytics, trade log, alerts, and market data. Slack integration delivering intelligence where the team works. The system is running — it just isn't touching real capital yet.

Month 4: Communications — LP & Content Pipeline

What we build: Automated LP letter generation, thought leadership content (2-3 pieces per week), meeting intelligence for fundraising.

Why it depends on Months 1-3:

• LP letters require performance data from the trading system (Month 3) and macro context from the research corpus (Month 1).
• Thought leadership needs the team's investment thesis as input — this emerges from trading sessions, not from a prompt.
• Meeting intelligence (prospect research, follow-up drafts) builds on the knowledge graph that has been accumulating since Month 1.

What the team gets at end of Month 4: Automated first drafts of LP letters. 2-3 thought leadership pieces per week (dual-language capability if desired). Pre-meeting research briefs and post-meeting follow-up drafts for fundraising, LP, and distribution partner meetings.

Month 5: Intelligence — Scenario Modeling & Proprietary Signals

What we build: Overnight scenario analysis lab. Proprietary signal generator on Viska infrastructure.

Why it depends on Months 1-4:

• Scenario modeling requires 4 months of operational data to calibrate — what scenarios are relevant, what portfolio impacts are realistic, what probability ranges are meaningful for Viska's specific holdings.
• Signal generation requires the backtesting framework from Month 3 and the macro data from Month 1. Signals are backtested against historical data, walk-forward tested, and validated before deployment.
• The knowledge synthesiser (weekly reflection processing) needs 8+ weeks of operational data to identify patterns worth codifying.

What the team gets at end of Month 5: Overnight scenario reports with probability assessments and portfolio impact projections. Proprietary signals that emerge from Viska's own framework — not generic indicators available to everyone.

Month 6: Handoff — Distribution Intelligence & Ownership Transfer

What we build: Partner surveillance (distribution partner monitoring), complete documentation, team training, and full ownership transfer.

What "handoff" means concretely:

After Month 6, Viska operates the system independently. Nothing is locked in — no proprietary platform, no SaaS dependency, no vendor that must remain involved. The intelligence department belongs to the fund.

The Harness Engineering Thesis

Why does orchestration matter more than the AI model itself?

Stanford and Tsinghua research (March 2026) demonstrated that the same AI model delivers a 6x performance difference depending on how it is orchestrated. The model is a commodity — the same Claude, GPT, or Gemini is available to every fund. What determines performance is the harness: how agents are structured, what context they receive, how they coordinate, what boundaries they operate within.

Three principles guide every decision we make:

1. Subtract Before You Add

One company removed 80% of its available tools and saw better results. Why? Because an agent with 43 tools spends most of its reasoning capacity deciding which tool to use, not using it well. Our agents have focused, curated toolsets:

When we evaluated the TradingView MCP integration, we found it offered 30 tools. We adopted 8 and rejected 22 — sentiment analysis that was just keyword counting, news feeds redundant with our existing pipeline, exchange data outside our scope. Selective adoption is a discipline, not a limitation.

2. Build the Layer That Survives

AI models are replaced every few months. Claude 3 → Claude 3.5 → Claude 4 → Claude 4.5 — each generation is more capable than the last. Any investment in a specific model's quirks depreciates rapidly.

The orchestration layer — how agents are structured, what data flows between them, how risk is enforced, how the team interacts with the system — transfers across model generations. When a better model ships, we swap it in and the entire system improves. The harness compounds; the model depreciates.

This is why we invest in architecture (agent boundaries, risk schemas, knowledge graphs, portal design) rather than in prompt tricks for a specific model version.

3. Govern First, Optimize Second

It is tempting to optimize for speed: faster analysis, faster execution, more trades, more signals. But speed without governance produces faster mistakes.

Every component has governance before optimization:

• Credential isolation before API integration
• Risk enforcement before order execution
• Audit trails before automated decisions
• Scope boundaries before tool access
• Human approval gates before any consequential action

The February 2026 supply-chain attack affected a platform that optimized for capability before governance. 21,000 instances with no authentication. 1,184 malicious skills in the marketplace. When the attack hit, there was no audit trail to determine which accounts were compromised, no scope boundaries to limit damage, no credential isolation to protect wallets.

We build governance into the foundation so that when (not if) something unexpected happens, the blast radius is contained and the audit trail is complete.

Capability Progression

What the team can do at each milestone:

+ = introduced that month. Once introduced, capabilities persist through all subsequent months.

Contact

Bóas Karl Svavarsson

hello@galdr.gg