Senior Android engineer and AI Native Engineer with 20+ years in software and 12+ on Android. Key contributor to two high-traffic Chilean banking apps with more than a million installs each. Builds on-device AI in the open with Knotwork, a local-first agent for Android, developed spec-first with coding agents under human review. Has shipped companion apps for BLE and Wi-Fi hardware, keeps crash-free sessions above 99% across releases, and has worked in distributed teams for 15 years.
His wordSource: his CV, May 2026
Case filesSelected projects
Knotwork
Verifiedchecked 2026-11-20
Open-source AI agent that runs on the phone
The model runs on the phone through LiteRT-LM, and every request goes through a pipeline you can read and change. By default a risky action waits for your confirmation.
Pipeline editor with 14 node types, in the app and as a single offline HTML file.
Each run records its seed, sampler, versions and the SHA-256 of every model file. A finished run can be repeated on the device and compared byte for byte.
SQLCipher database, API keys sealed under an Android Keystore key, and a build check that keeps exports off the network.
MCP tools with allowlists enforced where the tool runs. Interface, bundled pipelines and help in English, Spanish and Russian.
A destructive tool call waits for approvalPipeline editorTools with risk levels and MCP servers
lot-map
Verifiedchecked 2026-11-20
A live map of what Claude Code is doing
A panel beside the chat: the agent in the centre, ten stations around it, and a pulse for every tool call, model step and hook. The idea and the decisions are his; the code is his agent's.Open-source mod for Claude Code: a live map of the agent's work, drawn beside the chat. Idea and decisions by Aleksei, code by his AI agent.
Published on 9 November 2026. Every change of his own is committed with the agent's co-author line. Changes go in through pull requests, and a check validates the plugin and runs its tests on each one.
Labels carry names only: no command lines, file contents, memory text or passwords. A reader's audit of the code found a case that broke this rule, and the fix went out the same day, with tests.
Outside contributors have 22 pull requests merged, each read and run first: among them a fix for Windows paths, a fix that kept query text off a label, and labels in two more languages.
It makes no network requests, reads no files and runs no processes, and a check on every pull request holds it to the list of what it calls. Labels in 7 languages.
Board with one seed selectedA matched pair scores pointsColor symbols turned on
AI Native EngineeringAI Native Engineering
Coding agents work across the whole SDLC. The engineer owns specification, architecture and release decisions.
How a change travels. Select a stage to see its practices.
Under every stage
agents worka person decidesKeeper's wordObserved from inside the process
Spec-Driven Development (SDD)
Each task starts with a written specification: affected classes, interfaces, tests and files. Implementation begins only after it is approved.
Observed from inside the process
AI-Driven Development (AIDD)
Agents carry research, implementation, tests, review and refactoring through pull requests. Official documentation is checked before any library or API is used.
18 of the last 30 commits name an AI co-author. Merged pull requests: 619.
Human-in-the-loop gates
Stage gates at plan approval, design handoff, pull-request review and release. Irreversible actions require explicit human approval.
Observed from inside the process
Context engineering
An agent manifest as the source of truth, path-scoped rules, a generated code map, and long-term memory retrieved at session start.
Observed from inside the process
Multi-agent orchestration
Subagents for parallel research and independent code review, reusable commands and skills, tools connected over MCP.
Observed from inside the process
Guardrails as code
Pre-execution hooks on shell, git and pull-request operations. An LLM gate rejects commits and pull requests whose claims name no evidence.
Observed from inside the process
Agentic code review
The agent re-reads the real diff as a reviewer against a checklist. Each defect gets a regression test, and the pass repeats until it finds nothing new. A human review follows.
Observed from inside the process
CI quality gates
Detekt, Ktlint, Android Lint, a 75% Kover coverage floor, Konsist architecture tests, Roborazzi screenshot regression and instrumented tests.
Golden screenshots: 799. CI workflows: 4.
ADR and docs as code
Architecture Decision Records for every significant decision. Documentation changes ship in the same pull request, with link checks in CI.
Observed from inside the process
Release engineering and supply chain
Phase integration branches keep main releasable. Dependency verification with PGP and SHA-256, CI actions pinned to commit SHAs, secret scanning.