Viska × boas.dev / AI Intelligence Department

Proposal · Private · CF Access

Prepared for Guðlaugur Steinarr Gíslason · CIO

Engagement 6-month build & handoff

Submitted by boas.dev

Doc version v1.0 · 2026-04-19

State of the Art — Agent Harnesses

Control. Every output inspected.
Throughput. Capped by operator attention.
Audit. Conversational, hard to replay.

Control. Signed at the gates.
Throughput. Agents in parallel, humans review artifacts.
Audit. Every step a JSON artifact, replayable.

Control. Envelope, not gate-by-gate.
Throughput. Scales with hardware.
Audit. Mandatory, reversibility required.

A harness is the scaffolding around an AI agent — typed inputs, typed outputs, explicit gates, and a full audit trail of every decision. The harness is what turns a clever responder into something a fund can actually run.

Stage 1 is where most of the industry still lives. Human writes, AI answers, human reads, re-prompts. Useful. But the memory lives in one operator's head and the work does not survive the session.

Stage 2 is where boas.dev operates today. The Viska trading war room is a harness. Signal validation, risk check, execution — each a separate agent with a defined contract. The operator signs the order. The machine never acts alone.

Stage 3 is the destination, not the starting point. Agents execute within an envelope earned through track record. Reversibility is the price of admission. Viska moves there one audited cycle at a time.

Behind This Brief 01 / 04

Investigations

Nine domains. Every tool evaluated before adoption.

Research → Evaluate → Decide → Build

What we investigated

01n8n node capabilities — R2, SQLite, notifications, FinBERT. R2 adopted as static data store, ntfy for alerts.Adopted
02External API integration — Grok, Fed, ECB, Gemma. Responses API + grok-4-1-fast. Fed + ECB pipelines built.Adopted
03MCP & CLI tool audit — 11 services, 4 new MCP servers. FRED + QuantOracle close the quant-macro layer.Adopted
04TradingView MCP evaluation — 2 projects. One declined on ToS. One selective adoption for backtesting.Selective
05MCP vs CLI architecture — 25 runs. CLI 100% reliable, MCP 72%. CLI-first strategy adopted.Adopted
06Karpathy wiki pattern. Three-layer index retrieval beats RAG at ~400K words. No vector DB overhead.Adopted
07NotebookLM pipeline comparison. Claude pipeline wins 99% vs 60% on transcript success. Not integrated.Declined
08Security evaluation — supply-chain audit. Mastra pinned to SHA. 3 remaining vectors tracked with enforcement.Resolved
09Mastra core upgrade analysis — 18 releases. 2 breaking changes, neither affects our usage. Direct jump committed.Complete

Agent Roles

Every fund function mapped. 4 built, 10 planned, 2 new.

Roster pending Gulli's review

Role roster — fund operations

01Trading Operator — session coordination, order approval, thesisBuilt
02Technical Analyst — charts, patterns, trade idea generationBuilt
03Risk Manager — position sizing, exposure limits, drawdownBuilt
04Execution Agent — order assembly, validation, Alpaca submissionBuilt
05Macro Researcher — central banks, fiscal policy, ratesPhase 1D
06Crypto Analyst — on-chain metrics, exchange flows, DeFiPlanned
07Sentiment Analyst — NLP sentiment from news + socialPhase 1D
08Geopolitical Scanner — filings, sanctions, elections, policyPlanned
09Quant Strategist — backtesting, signal optimization, factorsPhase 1D
10Portfolio Analyst — cross-fund exposure, correlationNew
11LP Communications — letters, attribution, quarterly reportsMonth 4
12Content Producer — thought leadership, Substack, LinkedInMonth 4
13Meeting Intelligence — prep + follow-up, family offices, LPsMonth 4
14Partner Intelligence — distribution-partner surveillanceMonth 6
15Knowledge Synthesiser — weekly reflection, skill libraryPhase 1D
16Compliance Monitor — AIFM thresholds, position limitsNew

Pipelines Built

5 running on cron, 2 ready to schedule. Self-hosted, auditable.

n8n · auditable · self-hosted

Automation pipelines

01Watchdog — position monitoring, L1/L2/L3 alerts, every 5 min (08:00–20:00 ET)Testing
02Market Data — Alpaca + FRED → JSON → R2. Scheduled.Testing
03Trade Log — agent trade files → SQLite + R2. File trigger.Testing
04Signal Intake — TradingView webhook → validate → enrich → SQLite → R2 → notifyTesting
05Weekly Report — QuantStats metrics, P&L attribution, narrative. Fridays.Testing
06Grok Narrative — AI market narrative from macro + price data. Scheduled.Built
07CB Documents — Fed FOMC + ECB rate processing. Scheduled.Built

06mo

Engagement Plan

Six milestones. Reprioritize at any phase boundary.

Phase-boundary reprioritization

Monthly milestones

M1Macro Research Intelligence — full corpus queryable, daily brief, MCP workspace accessBuilt
M2Multi-Agent Surveillance — 3 autonomous agents (crypto, macro sentiment, geo-risk) 24/7 + alert pipelineNext
M3Trading War Room — paper trading, 4 agents, 7 pipelines, risk enforcement, 7-page dashboardQueued
M4Investor Communications — LP letter pipeline, thought leadership (2–3/wk), meeting intelQueued
M5Scenario Modeling & Signals — overnight scenario lab, proprietary signal generator on Viska infraQueued
M6Distribution Intelligence & Handoff — partner surveillance, codebase handoff, documentation, trainingQueued

Navigation

Use keyboard shortcuts or click a section to open.

1–5Open section by number

←→Page between sections

ESCReturn to index

?Keyboard reference