Portfolio

Selected engineering proof: what I built, the systems underneath, and the tradeoffs that mattered. Raintree product pages live at raintree.technology.

Agent Safety & Evals

Context & Developer Tools

Data & Intelligence Systems

Product SaaS

Systems & Low-Level

Music & Creative Systems

Research graph / personal case study

Investment Research Graph

local intelligence pipeline

What I built

Private cross-asset research graph with local data pipelines, MCP tools, and owner-approved mutations.

Stack

Python 3.12 / FastAPI / FastMCP / SQLite / sqlite-vec / DuckDB / Parquet / NetworkX

Runtime

One pure-Python tool layer backs MCP stdio, an HTTP sidecar, CLI commands, and the local UI cockpit.
Nightly refresh runs prices, macro, financials, analyst inputs, graph rebuilds, embeddings, proposal generation, review, and brief delivery as isolated phases.
LLM proposer and reviewer tiers write candidate relationships to staging before owner curation promotes them to pending actions.

Data model

SQLite stores nodes, edges, pending actions, staging proposals, proposal logs, security identifiers, financials, macro series, and analyst metadata.
sqlite-vec stores 1024-dimensional node embeddings for semantic lookup and entity reconciliation.
DuckDB and Parquet hold EOD prices, returns, macro series, financials, briefs, and data-quality warnings.

Techniques

Owner-only curation CLI is the default approval boundary; programmatic approval is disabled unless explicitly enabled.
Sidecar binds to loopback by default and requires bearer auth before any non-loopback exposure.
Vendor failures are recorded per pipeline phase without corrupting later graph or brief stages.
Security-master identifiers, ticker foreign keys, quarantines, and pruning migrations keep market-data joins auditable.

Verification

Tests cover agent proposal flows, brief delivery, DB durability, financial data quality, graph relationships, signal relationships, MCP tools, sidecar auth, and ingestion.
Launchd schedules separate refresh and brief-delivery jobs so research generation and notification delivery can fail independently.
Approval gates keep graph mutation separate from LLM proposal and review.
The public portfolio version is sanitized to describe architecture without exposing personal holdings or raw local data.

Portfolio

Selected engineering proof: what I built, the systems underneath, and the tradeoffs that mattered. Raintree product pages live at raintree.technology.

Agent Safety & Evals

Context & Developer Tools

Data & Intelligence Systems

Product SaaS

Systems & Low-Level

Music & Creative Systems

Research graph / personal case study

Investment Research Graph

local intelligence pipeline

What I built

Private cross-asset research graph with local data pipelines, MCP tools, and owner-approved mutations.

Stack

Python 3.12 / FastAPI / FastMCP / SQLite / sqlite-vec / DuckDB / Parquet / NetworkX

Runtime

One pure-Python tool layer backs MCP stdio, an HTTP sidecar, CLI commands, and the local UI cockpit.
Nightly refresh runs prices, macro, financials, analyst inputs, graph rebuilds, embeddings, proposal generation, review, and brief delivery as isolated phases.
LLM proposer and reviewer tiers write candidate relationships to staging before owner curation promotes them to pending actions.

Data model

SQLite stores nodes, edges, pending actions, staging proposals, proposal logs, security identifiers, financials, macro series, and analyst metadata.
sqlite-vec stores 1024-dimensional node embeddings for semantic lookup and entity reconciliation.
DuckDB and Parquet hold EOD prices, returns, macro series, financials, briefs, and data-quality warnings.

Techniques

Owner-only curation CLI is the default approval boundary; programmatic approval is disabled unless explicitly enabled.
Sidecar binds to loopback by default and requires bearer auth before any non-loopback exposure.
Vendor failures are recorded per pipeline phase without corrupting later graph or brief stages.
Security-master identifiers, ticker foreign keys, quarantines, and pruning migrations keep market-data joins auditable.

Verification

Tests cover agent proposal flows, brief delivery, DB durability, financial data quality, graph relationships, signal relationships, MCP tools, sidecar auth, and ingestion.
Launchd schedules separate refresh and brief-delivery jobs so research generation and notification delivery can fail independently.
Approval gates keep graph mutation separate from LLM proposal and review.
The public portfolio version is sanitized to describe architecture without exposing personal holdings or raw local data.

Selected projects

What I built

Stack

Selected projects

What I built

Stack