Project

Blockwise Intelligence

An autonomous AI-native quantitative finance research lab. Co-built with Finn Westerink. Fresh-start brand as of May 2026 — building intentionally from the one-pager, with the Master OS as the operational substrate.

Role Co-Founder, Architect & Operator

Co-founder Finn Westerink

Status Early-stage / active build

Since May 2026 (fresh start)

What Blockwise Intelligence is

Blockwise Intelligence is an AI-native quantitative finance research lab. The core thesis: most quant research infrastructure is built for teams of analysts — expensive to run, slow to iterate, and dependent on proprietary data pipelines that take months to stand up. AI-native infrastructure changes that equation.

With autonomous agents handling research generation, backtesting orchestration, and signal evaluation, one or two operators can run a research pipeline that previously required a full team. Blockwise is being built from that premise up — not retrofitting AI into legacy quant workflows, but designing the architecture from scratch for an agentic-first world.

The brand is fresh as of May 2026. Prior exploratory work runs preserved but deprecated. Building with intention from a clean one-pager.

The isolation architecture

Blockwise Intelligence runs inside the Master OS alongside six other business domains. The design constraint is strict: financial research context must never bleed into other domains. The solution is defense-in-depth — four independent layers, any of which can fail without breaching the boundary. All four must fail simultaneously for cross-contamination to occur.

Layer 1 Config deny-list

.claude/settings.json deny-list blocks file reads at the Claude Code configuration layer — before any AI tool call is processed.

Layer 2 Advisory rules

.claude/rules/blockwise.md is loaded for every session. Explicit boundary instructions are always in context.

Layer 3 PreToolUse runtime hook

boundary-guard.json intercepts every Bash, Read, Write, and Edit call. If the target matches Blockwise silo paths, the hook blocks at the runtime level — independent of model intent. This is the deterministic layer.

Layer 4 Python runtime guard

scripts/safety_guard.py provides a Python-level check at every Bash invocation across all ticket executors.

What this signals

Security discipline at the architecture layer. Treating financial research as a distinct security domain — enforced by infrastructure, not convention. Same instinct that makes a good security or systems engineer.

AI-native from the ground up. Not "add AI to existing quant stack." Designing the research pipeline for autonomous agents as first-class operators.

Operational range. Running a quant research system concurrently with five other active domains — keeping all of them cleanly separated — is an operational challenge most engineers haven't faced. The system handles it.