Inside the AI Lab · Updated 2026-09-09
Tech stack
A working environment for investment banking, research and publishing. Local machines, frontier models and specialist agents share the work, with curated knowledge and human judgement connecting the layers.
Explore the stackRead the AI Lab →
Local-first, with cloud capacity where it helps.
Compute fleet
- MacBook Pro
- The primary interactive node: M4, 64 GB unified memory and 1 TB storage in the source inventory. Hermes and local models support hands-on analysis and experimentation.
- Mac mini
- An always-on M4 background node with 16 GB unified memory, hosting Samantha on OpenClaw.
- Ubuntu ThinkPad
- A lightweight local node for the Stöbä/Hermes environment and home-automation mesh control.
- Windows ThinkPad
- A secondary Windows 11 Pro machine with 32 GB memory, listed without an active agent workload.
- Planned desktop inference node
- The source lists a Mac Studio order aimed at larger local inference and an expanded Samantha environment. This is a plan, not a claim of a deployed machine.
- Cloud workspace
- Cursor Team and Grok Bot complement the local fleet with repository work and specialist cloud agents. Tailscale and SSH connect the working environment.
Choose the model for the task, not the other way around.
Intelligence layer
- Anthropic / Claude
- Long-context reasoning, document synthesis, negotiation preparation, investment memos and coding through Claude Code and desktop workflows.
- OpenAI / GPT and Codex
- Multimodal reasoning, visual context, agentic coding, data pipelines and MCP server development.
- Google / Gemini
- Multimodal analysis, long-video understanding, visual work, market scans and code evaluation.
- DeepSeek
- An alternative reasoning engine and a fallback option in Samantha’s model-routing setup.
- Local models
- The source lists Gemma, Qwen and Llama experiments. Local inference is useful when data boundaries, offline operation or control of the runtime matter.
Different roles, explicit boundaries.
Agents & orchestration
- Samantha / OpenClaw
- The persistent background agent on the Mac mini: context, coordination and a continuing record of work.
- Hermes
- The interactive agent environment on the MacBook Pro, combining local models with frontier reasoning.
- Stöbä
- A specialist local agent environment associated with the Ubuntu ThinkPad and investment-banking workflows.
- Grok Bot
- Specialist cloud-agent work spanning sites, episodes, YouTube, the work graph and sparring.
- Office & document workflows
- Claude Cowork, Excel, Word and PowerPoint support modelling, valuation, drafting, contract review, board memos and presentations.
- Building & research surfaces
- Codex, Cursor, Lovable and Perplexity Comet cover software, app prototyping and live research.
- Generative media
- Fal.ai, Seedance and Gemini-based workflows form the source dashboard’s media-generation and multimodal experimentation layer.
Useful intelligence starts with useful context.
Knowledge & data
- Market and company intelligence
- Inven for deal and company research; Inderes MCP for Nordic equity research, estimates and transcripts; S&P Global for financial data, screening and valuation work.
- Legal research
- laki.ai connects Finnish legislation and court decisions to research workflows. The source also lists oik.ai among the connectors.
- Relationship and collaboration systems
- HubSpot provides CRM and pipeline context. Microsoft 365, Outlook, SharePoint and Teams support the collaboration layer.
- Google workspace connectors
- Gmail, Google Drive and Google Calendar sit alongside OneDrive and Slack in the connector inventory.
- Publishing knowledge
- The Neuvottelija podcast and article corpus support episode research, negotiation preparation and source-grounded publishing.
- Memory & engineering
- Memori and curated memory provide cross-session context. GitHub, filesystem tools, PDF tools and Supabase support code, documents and application data.
- Samantha Grid
- The source describes a multi-domain workspace for analysis, discussion and bounded code execution.
Repeatable methods, not a fresh prompt every time.
Reusable skills
- Translink style
- House-style guidance for investment memoranda, decks and business writing.
- Finnish legal review
- Contract review, clause extraction and risk-flagging workflows, with professional judgement retained.
- Second Brain
- Layered, curated memory for carrying useful context between sessions.
- Multisource Daily Brief
- Python and CLI-assisted collection and curation of research inputs.
- Response workflows
- Email, WhatsApp, Telegram and Discord workflows with differentiated trust and approval levels.
- Research & episode production
- Inven and Inderes workflows, EXA research with citations, and YouTube summaries with timestamps.
- Suomen Kieli & Skill Creator
- Finnish language and style checks, alongside a method for building and maintaining new skills.
Readable by people. Discoverable by search and agents.
Web & publishing stack
- Neuvottelija.com
- Astro, Git and Cloudflare Pages. Static HTML, Markdown mirrors, structured data and sitemaps support the international publication and AI Lab.
- Neuvottelija.fi
- React/Vite on Cloudflare Pages, with Supabase-backed features and build-generated public content. Finnish guests, conversations, transcripts and services.
- DCM Capital & Translink Finland
- The source identifies WordPress as the publishing platform for dcmcapital.fi and translinkcf.fi.
- Discovery surfaces
- Canonical URLs, source citations, RSS, sitemaps, llms.txt and machine-readable indexes. These are maintained publishing interfaces, not a promise of a particular search ranking.
Capability does not remove the need for judgement.
Security & operating principles
- Data boundaries
- Minimise the information passed to a model, distinguish public research from confidential mandate material, and use local processing where appropriate.
- Controlled connectivity
- Private connectivity, SSH and scoped tools support access boundaries. A connector’s presence is not permission to share every record it can reach.
- Human review
- Legal, financial and externally consequential work needs appropriate review and action-specific approval. Model selection is not a substitute for accountability.
- Persistence & recovery
- Time Machine, application storage and curated memory serve different roles: backup, system of record and working context should not be confused.
- Compliance is workflow-specific
- Privacy obligations, confidentiality and applicable AI rules need assessment for the actual data and use case. No blanket GDPR, EU AI Act or NIS2 certification is asserted here.
From architecture to practice
Watch the agents at work.
A Neuvottelija Sisäpiiri demonstration introduces Samantha and the agent workflow behind the environment.
No video or third-party player loads until you choose to visit YouTube.
About this inventory
An editorial snapshot adapted from Sami’s AI Environment dashboard, not live telemetry. Hardware plans, model choices and integrations evolve. Listings describe the environment, not guaranteed availability, vendor specifications or compliance certification.
Adapted from Sami Miettinen AI Environment. For the evolving experiments and implementation notes, explore AI Lab. For a business discussion, contact Sami.


