I build agent systems that plan, use tools, verify their work, and operate safely around code and data. Below are those systems, followed by selected ML, data, and research work.
I. Agent Systems
spec
goal
race / ratchet
reviewed PR
Versioned agent workflows for long-horizon execution, multi-model review, measured optimization, and agent-tool hygiene.
- Long-haul loops turn a request into a verifiable goal and keep only measured wins
- Dual-haul races independent Claude and Codex worktrees under the same check
- Pair workflows add bounded consultation, optimization, review, and test gates
audit
agent teams
TDD gate
merge
Multi-agent system that audits a codebase, finds high-value improvements, and ships only mechanically verified changes.
- High-planner dispatches scoped work to isolated, disposable worktrees
- Non-LLM verifier proves red → green and re-runs the full test suite
- PyPI CLI, HTML reports, and no-push/no-bulk-stage safety boundaries
Agent-driven hyperparameter optimization with Optuna — Claude Code adapts the search strategy round-by-round.
- Agent diagnoses convergence and parameter importance through MCP
- Optuna keeps every round deterministic across three boosting backends
- Guardrailed decisions and self-contained campaign reports
Foundation layer for agents over Postgres — agent-curated materialized views behind a two-role access boundary.
- Setup and runtime MCP servers; agents never touch base tables
- Access boundary is enforced inside Postgres, not the app layer
- SQL lineage, refresh history, and a static operations dashboard
Agent-native CLI that discovers compute-heavy pandas operations and validates their Polars conversions.
- Ranks conversion targets by estimated impact using AST analysis
- Checks equivalent output before benchmarking time and memory
- Returns structured JSON designed for autonomous optimization loops
II. Applied ML & Data Systems
Integrated drone vision system with trajectory planning, object detection, and visual SLAM.
End-to-end pipeline for multi-class segmentation and classification on prostate MRI data.
- Config-driven modular architecture with layered YAML composition
- HPC/SLURM sweeps with automatic model promotion
- Research runner orchestrating data import through inference
Event-driven inventory sync replacing daily batch pulls with near-real-time Shopify webhook updates.
- Shopify webhooks → FastAPI → Kafka (KRaft) → PostgreSQL
- Idempotent and order-safe with partition-keyed messages
- Containerized with a mock webhook generator for local testing
Plain-text, git-tracked personal portfolio manager — check, visualize, and rebalance from terminal or dashboard.
- Drift-to-target in one command with EUR-native FX conversion
- Trades live in CSV, targets in YAML, and git is the audit trail
- Streamlit dashboard with inline editing in about 500 lines of Python
III. Publications
Deep Learning-Based Semiautomatic Generation of HD maps from Aerial Imagery
