Mímir — Architecture, Capabilities & Roadmap The fund's AI research analyst: what we've built, how it runs, and where it's going.

01Executive Summary

Mímir is Viska's AI research analyst. A simple Slack surface sits on a production system: Mastra agent framework, Supabase vector store, a separate n8n ingestion pipeline, ~1,200 passing tests.

02What Mastra Is

Mastra (@mastra/core) is an open-source TypeScript framework for production AI agents. It provides the scaffolding we'd otherwise hand-build:

Model-agnostic. Mímir uses OpenAI today — gpt-5.2 (reasoning), gpt-4o-mini (lightweight synthesis), text-embedding-3-small (1536-dim embeddings) — but is not locked to it.

03How Mímir Is Deployed

Runtime: Node 24, Fastify HTTP, Slack Bolt in Socket Mode (outbound WebSocket — no public inbound URL). Boots observability → advisory connectivity checks → routes → Slack. Every dependency degrades gracefully.

Deploy pipeline (Railway)

feature/* ──local──► develop ──auto──► mimir-staging   (Dev Slack, #viska-agent-dev)
                                          │ verify
                       main ◄── PR merge ─┘
                        │
                  auto-deploy (branch-tracked)
                        ▼
                     mimir  (PROD ← Viska Capital Slack)

• Project viska-mimir-agent · services mimir (prod←main) and mimir-staging (←develop).
• Prod is branch-tracked: merging develop→main is the deploy. Local railway up to prod is forbidden.
• 3-layer gate: pre-commit (tests) → pre-push (blocks main unless flagged + clean + green) → deploy:check.

The Mastra-service vs n8n-pipeline boundary

Live DB schema is owned by the ViskaDB repo's migrations. Embedding model + dimension are locked to match ingestion.

04Total Architecture Map

┌──────────────────────────── SURFACES ────────────────────────────┐
│  Slack (@mention, /mimir, App Home)        viska.gg /chat (next)   │
└──────────────┬────────────────────────────────────┬──────────────┘
               │ Bolt Socket Mode                    │ HTTPS/SSE (CF Worker — deferred)
               ▼                                      ▼
┌──────────────────── MÍMIR  (Mastra service · Railway) ────────────────────┐
│  Fastify + Slack Bolt                                                     │
│   mimirAgent  (gpt-5.2, maxSteps 5)                                       │
│    ├─ 20 tools (thin adapters → shared core)                              │
│    │   corpus search · live market/crypto · reports · prefs · Icelandic   │
│    │   sector sentiment · web search · custom HTML report                 │
│    └─ MODULE SYSTEM  (registry → manifest SSOT → 3 surfaces + self-explain)│
│          └─ last30days  (BASE: reddit/hn/github/web ·                     │
│                          FINANCE: edgar/gdelt/polymarket/+dormant tier-2) │
│   Memory: working (per user) + thread        RequestContext (per call)    │
│   RAG: embed → hybrid_search_v3 (vec .7 + kw .3, RRF) → synthesise        │
│   Face layer (built, not yet consumed): face-jwt · face-resolver          │
└──────┬──────────────────────────┬───────────────────────────┬────────────┘
       ▼                          ▼                           ▼
  ┌──────────┐        ┌────────────────────────┐     ┌──────────────────┐
  │ OpenAI   │        │ Supabase (pgvector+PG)  │     │ n8n pipeline      │
  │ LLM/embed│        │ embeddings · mimir_*    │     │ ingest · reports  │
  └──────────┘        │ viska_tickers · memory  │     │ error triage      │
  ┌──────────┐        │ sector scores · errors  │     └────────┬─────────┘
  │ Binance/ │        └────────────────────────┘    Dropbox (docs) ◄┘
  │ DeFi/    │
  │ Perplex/ │
  │ Miðeind  │
  └──────────┘

Four layers: external data sources → durable state (Supabase + n8n ingest) → the Mímir runtime (Mastra + tools + modules + RAG) → user surfaces. The face layer is the seam that lets one agent safely serve many surfaces.

05How We Build Skills Now — the Module System

Capabilities used to be 20 flat, hand-registered, inconsistently-named tools. No concept of "a skill," no slash commands, no guide, no self-explanation. The module system (live on prod) replaced that with a rail.

A module is a folder src/modules/<id>/ declaring everything through one manifest (the single source of truth). A central registry wires that one manifest into three user surfaces plus a docs plane:

• Natural language — the agent gets the tool and decides when to use it.
• Slash command — /mimir <skill> <args> (one umbrella command; Slack configured once, forever).
• App Home — a Skills card + guide modal on the dashboard.
• Self-explain — an explain_skill meta-tool reads the same help text.

Net for Viska: adding an analyst capability is now a bounded, repeatable task — write a module, get three surfaces + a guide + tests for free.

06Face Agents — "One Agent, Many Faces"

One Mímir, multiple surfaces. A face controls what data the agent may see and do — enforced in three layers of increasing trust.

1. Capability — which tools are mounted this request.
2. Data — which DB rows it can read/write (Supabase row-level security via a face-scoped, short-lived signed token).
3. Memory — which conversation history it recalls (per-shard).

Status — audited against live code

The honest gaps (work remaining)

Order: G2 → G1 → G4 (ViskaDB) → viska.gg edge-auth (ViskaFront) → G5 → G6. Feeds the larger per-user tenancy epic (board #7), gated on ViskaDB identity-unification (Slack text-IDs vs Supabase UUIDs) + OpSec JWT-revoke. Operator-review standby — not started unprompted.

07What This Unlocks for the Fund

Viska is a $15.5M macro fund (21 live positions — metals/mining, energy, grid/power, semis) run against a 19-theme × 16-quarter strategic grid (Gold 22.1% NAV, Copper 15%, Energy Infrastructure 14.6%, Aluminium 11.6%…). The edge is regime-aware theme rotation — so the scarcest resource is early, synthesised signal on theme shifts, not raw data volume.

Mímir is the interface that turns the fund's data estate into that signal, conversationally:

• Research corpus on tap — embedded analyst research, hybrid-searchable, cited back to source docs. Live.
• Live market awareness — crypto/derivatives/DeFi/sentiment, plus finance-aware last30days (filings, news, prediction-market odds, fund-ticker detection).
• The fund's own framework — reasons in Viska's themes, tickers, and followed voices (Alden, Gromen, Howell, Wang, Werner, Hayes) — not generic.
• A path to the full book — trading-intelligence read views (#8) and the quant conviction engine (#9, a per-broker × per-theme council, already writing live signal) plug into the same agent.
• Whole-team access, safely — the face/tenancy arc puts Mímir on the viska.gg portal, each user seeing only their own threads and permitted data.

08Work State — GitHub Boards (SSOT)

Shipped (30d): migration to Viska org · sharding foundation S1/S3/S4/S5/S6/S8/S9 · module system M1–M5 · last30days finance-signal · prod ship today.

Active blockers: Phase-C cutover (drifted Phase-A apply) · sharding S2/S7 (needs operator relaunch + CI-gate call) · Tier-2 finance keys (dormant) · face/tenancy build (greenlight-gated).

09Logical Next Steps

Now — low-risk, unblocks signal

Next — the operator's decision

Then — capability depth (mostly externally-gated)