Skip to content
← Work

Daniil Maximkin · Engineering notebook

Lab

Tools I use, prototypes I built and ideas I tested. I keep the method, the current stage and the limits together.

Internal tool

Shopify speed kit

Status
Built internal diagnostic tool
My role
Tool designer and engineer

I built a repeatable way to collect storefront evidence and turn it into a diagnosis. The scan looks for benchmark cloaking, duplicate scripts and render-blocking work. The AI layer ranks causes and records uncertainty.

  1. Collect evidence

    Deterministic audits and script scans. Public diagnosis does not require store access.

  2. Propose a fix

    Ranked causes, confidence, risk, reversibility and the checks each change needs.

  3. Apply with approval

    Theme changes follow an approved proposal and a backup. A model does not silently write to a store.

A lab score is evidence from a test run. I do not promise a score increase or treat it as a real-user measurement.

Local prototype

CCTV AI

Status
Local prototype; deployment not verified
My role
Prototype engineer

I built local scripts around camera archives: timelapse extraction, object detection, event clips and Telegram reporting. The project is designed for CPU processing on macOS, without a continuous live stream.

The scope is object events, not face recognition or identifying people. I do not publish camera frames, locations or claims about gate safety.

Demo / simulator

Digital twin

Status
Demo and simulator only
My role
Demo designer and engineer

I built an operator interface, a simulation core and a draft backend gateway to explore an equipment workflow. It is a way to discuss states and architecture before connecting real hardware.

  1. Operator interface

    A local browser UI for presenting the workflow.

  2. Simulation core

    Equipment states in a prototype, with state-machine and logging work still open.

  3. Gateway boundary

    A draft gateway. The hardware adapter remains unfinished; no live equipment control is established.

Research notebook

Information Gravity

Status
Paused experiment; recall improvement not validated
My role
Experiment author

I explored whether mutual-information links could help a model recover facts across gaps in its context. The experiment includes a demo and diagnostic scripts.

I do not publish the recall multiplier in the old README or link to its hosted demo. That claim is not established by the evidence I have.

Hypothesis → diagnosis

Question
Could a mutual-information bias help recover a distant fact?
What the diagnosis found
The matrix had no context-to-target links in the checked examples.
Limit
The diagnostic note describes a reduced-gap demo. It does not validate the original long-gap claim.

Local demo foundation

Warehouse PWA

Status
Executable local demo foundation; production not verified
My role
System designer and engineer

I built a foundation for scanning and inventory workflows: a PWA, API, typed contracts and a background worker. Inventory commands go through a gateway rather than a second editable stock balance.

The README describes local demo authentication and no enabled external catalogue provider. It does not establish a production warehouse deployment or a client rollout.

Public demos

Code you can run locally

Status
Public demos; synthetic data only
My role
Demo author and engineer

I built these demos with synthetic data so you can run the code and inspect the decisions.

AI-channel order capture
I label orders, build GA4 purchases and reconcile the week; ad-platform payloads are never sent. The order-capture demo repository on GitHub.
Metafield-configured discounts
I keep market thresholds and tag exclusions in a metafield and test the Discount Function locally, including its WASM build. The Discount Function demo repository on GitHub.
ChatGPT Ads measurement
I check consent on the browser and webhook routes and give them one purchase ID; delivery is a dry run by default. The conversion-sender demo repository on GitHub.
Reorder and purchase-order drafts
I calculate sales velocity and low-stock signals offline, then draft supplier orders for explicit local approval; no emails or stock changes follow. The reorder demo repository on GitHub.