Selected work
Public projects and research, with their purpose, technical choices and supporting evidence.
Research · completed retrieval evaluationRAG Retrieval Research
A Russian-language retrieval study with separate full-corpus and reduced-collection experiments. The case includes measured TEST results and a DEV ablation.
Read the case · Repository
Experimental CLI · alphaSnapKeep
A Python command-line tool for timestamped project snapshots and archive verification. It has not had an independent operational trial.
Read the case · Repository
AI engineering · learning experimentAI Text Helper
A Russian text API prototype with improve and summarize tasks, a BM25 correction lookup and local-model integration. Fine-tuning results are not claimed.
Read the case · Repository
Developer tools and agent workflows
Two public tools and a local skill for carrying project context between agent sessions.
Agent workflow · instruction setAgent Safe Flow
Five reusable instructions for specification, audit, implementation, review and release checks. This is not a software runtime.
Read the case · Repository
Developer tooling · v0.1.0Codebase Atlas
A portable skill for source-backed codebase walkthroughs and Mermaid diagrams. It does not include a diagram renderer.
Read the case · Repository
Agent continuity · published skillContext Waypoint
A Markdown convention and agent skill that help carry concise project state and handoffs between work sessions.
Read the case · GitHub repositoryThe repository includes the README, skill, templates, documentation and a project example. Context Waypoint is a Markdown convention and reusable skill. It is not a CLI or hosted platform.
Engineering practice
Examples from public projects include backend reliability, retrieval evaluation and developer tools.
Backend systems and reliability
Keep API boundaries clear, validate inputs, handle failures deliberately, and consider safe data changes, testing, and observability.
Retrieval evaluation
The RAG Retrieval Research repository contains saved search runs and Recall@K and nDCG metrics. DEV ablations are separate from TEST results.
Developer tools and agent workflows
The public tools cover careful file handling, source-based code explanations and staged checks.
How I work
- Frame the problem and constraints.
- Design boundaries and failure behavior.
- Implement a coherent change.
- Validate expected and failure paths.
- Consider operations and maintenance.
Research notes
These notes are limited to the RAG experiments and collections described in the case. They may not apply to other datasets.
The tested hybrids showed no nominally significant gain
Paired tests found no nominally significant improvement over the dense baselines for the tested fixed-weight hybrid methods.
Results from the reduced collection apply only to that collection
Positive TEST qrels and historical retrieval results informed the collection, so the test results are not independent of its construction.
Read the research case
About
I value work that can be checked: clear interfaces, explicit limits, and conclusions supported by tests or measurements.
The projects above include a text API prototype, retrieval research, an early CLI tool and developer tools.