I design data platforms and bring machine learning into production. My background spans cloud architecture, distributed data processing, and engineering leadership.
Current focus — AI agents, reliable automation, and source-backed content research.
- Data systems: collecting external data, preserving its source, and making it useful downstream.
- AI workflows: connecting research, generation, validation, and measurable feedback.
- Engineering practice: small changes, repeatable checks, and evidence of what actually ran.
- Data platform modernization: Kubernetes-native, event-driven AWS systems with streaming ingestion, lakehouse storage, GitOps, and data quality controls.
- Data transformation at scale: engineering and development leadership for pipelines processing approximately 50 TB per month in an automotive engagement at Exadel.
- Production ML: model delivery, distributed training, continuous retraining, and inference for enterprise banking at SoftServe.
| Project | What it demonstrates |
|---|---|
| Autonomous code generation sandbox | A disposable Python target for an agent workflow: scoped task → implementation → pytest → pull request. An experiment, not a production framework. |
| Research workflow smoke test | A small validation target for a research workflow. |
Small, reproducible experiments in agent memory, retrieval, and reliable AI systems. Each note separates what the code demonstrates from what remains unmeasured.
- PageIndex: success does not mean every requested page came back — five synthetic cases against the real local retrieval function; check returned page IDs and recover omissions before synthesizing.
- Paperclip’s $0 budget is not a stop button — six synthetic boundary cases against the real budget-status helper; disabled caps and heartbeat pause are separate controls.
- How four second places become first — an original rank-fusion experiment using Hindsight’s pinned implementation.
- The classifier’s 100% came from two missing labels — why an omitted class is not a zero-probability class; a synthetic fixture against the actual parser.
- Can you take the reasons with the code? — an offline audit for project memory.