EPTHE MECHANICAL COMEDIANPROJECT SHOWCASE

Elias Penkkimäki / Selected projects

Practical systems.
Curious experiments.

I design, maintain and orchestrate tools that put AI to work in tasks where it produces actual value instead of merely hype. I explore how it can serve our purposes in creation and play, while keeping us in control rather than at AI’s mercy.

Explore work projects Visit the hobby lab
Local inferenceBusiness automationCreative softwarePlayable experiments

01 / Professional practice

Work projects06

Tools for everyday operations, connected business data, and AI environments for customer organisations.

01 / Knowledge accessIn use

Local AI instruction bot

The right guide, straight from the source. A Slack bot that finds customer and process instructions and returns the complete text with a source link.

Keep the instructions intact.

Local inference helps locate the guide; the answer preserves its wording. Missing or ambiguous customer information leads to a clarification.

  • Slack
  • Local LLM
  • Airtable
  • DOCX
How it works

Users ask through direct messages, mentions, or the /ohje command. The bot distinguishes an unknown customer from a known customer with no guide, and marks draft process instructions as drafts.

A local administration interface imports customer documents and refreshes process instructions from Airtable. An explicit index refresh makes updates available to the bot.

02 / Business intelligenceEvaluated

BigQuery AI

Bringing subscriptions, customers, and invoices into one data model, then asking business questions in natural language.

92%Subscription-to-customer match rate
10/10Questions passed in the March 2026 evaluation
  • BigQuery
  • Dataform
  • Cloud Run
  • Python
Pipeline & evaluation

Cloud Run connectors ingest Airtable and Briox data. Dataform transforms raw JSON through staging and core models into business marts. Normalised names, ID mapping, and fallback matching connect records across systems.

BigQuery Conversational Analytics answers questions in Finnish about subscriptions, billing gaps, and changes. All ten questions in the 17 March 2026 test report passed table selection, SQL correctness, and answer clarity checks.

The evaluation supported using Conversational Analytics without building a custom query API. Revenue coverage remains limited to Also subscriptions with available pricing; unmatched customers remain a data-quality limitation. This result describes the tested question set.

03 / Customer environmentsCustomer work

Copilot for business

Setting up Microsoft Copilot for business and Copilot Studio pay-as-you-go environments for customers.

From product to working environment.

Customer environment setup spanning business Copilot use and the Copilot Studio PAYG environment for building agents.

  • Microsoft Copilot
  • Copilot Studio
  • PAYG
04 / AI administrationCustomer work

Claude in customer organisations

Managing Claude for customer organisations and managing integrations within their Microsoft environments.

AI as part of the organisation.

Administration and integration work connecting customer AI tooling with the environment their teams already use.

  • Claude
  • Microsoft environments
  • Integrations
05 / Earlier approachRetired

Local Hermes agent

A local instruction assistant with the same guide-finding role as the instruction bot, plus the ability to receive instructions to update the guides.

Why it was retired

The Local AI instruction bot proved easier to maintain and less prone to security risks. The more focused tool became the preferred approach.

  • Hermes
  • Local AI
  • Guide maintenance
06 / Team coordinationRetired

Local Hermes Coordinator

A Slack coordination agent that gathered standup replies, relayed tasks, routed blockers to responsible people, and produced daily summaries.

A bounded coordination role.

The project explored daily team communication with a dedicated Hermes profile and tested restrictions around the agent’s work role.

  • Hermes
  • Slack
  • Workflow automation

02 / The personal lab

Hobby projects04

Experiments that connect hardware, storytelling, sound, and play. Built to explore an idea all the way to working software.

01 / Local AI infrastructureFull case study

WardenDual-GPU control plane

Two GPU vendors. One shared memory budget. A control plane that gives applications time-bounded leases on local inference, supervises model runtimes, and explains capacity refusals.

  • Python / FastAPI
  • CUDA + ROCm
  • llama.cpp
Explore the Warden case study Explore the Writer case study
A shared workstation, coordinated.
WardenSUPERVISE · BUDGET · ROUTE
NVIDIA / CUDARTX 509032 GiB
AMD / ROCmRadeon Pro W780048 GiB

80 GiB VRAM across two vendors

02 / Creative tools

AI-Home-Lab-Writer

A writing workspace designed to make AI useful as an assistant, organiser, and linker of project knowledge.

Keep the writer in control.

The goal is to simplify the writing process, make project knowledge easier to maintain, and surface conflicts for review. Warden and multiple inference fallbacks provide the model infrastructure.

  • Writing workflows
  • Warden
  • Knowledge organisation
Read the Writer case study
03 / Audio tools

LitListener

A Rust and Tauri desktop project combining OpenAI Whisper transcription with a workspace for listening to and analysing audio.

Keep the original. Compare the details.

Waveform playback and original/denoised audio comparison support careful listening. Transcripts stay linked to the original recording, with audio records stored in SQLite.

  • Rust
  • Tauri
  • Whisper
  • SQLite
Project background

The project moved from a Java-based backend to Rust and Tauri, retaining a web-based interface. Audio import, playback, processing, and transcript management come together in a desktop application.

04 / Browser games

Lure of the Abyss

A fishing-survival game where a small language model becomes the sea god, answering prayers and giving the crew a voice in the player’s browser.

The model is part of the game.

SmolLM2 runs in a Web Worker through Transformers.js and WebGPU. Players can also choose a complete non-AI mode with ritual buttons and written dialogue.

  • JavaScript / Canvas
  • WebGPU
  • SmolLM2
Play Lure of the Abyss

Source code & access

Interested in a project?

GitHub access cannot be provided for work projects. For hobby projects, please send an access request on LinkedIn.

Contact me on LinkedIn