Trusted Insights.
Decisive Operations.

Project Leo — Underwriting Intelligence

An AI assistant that autonomously ingests and cross-references public healthcare and population datasets with proprietary real estate data. It synthesizes a massive vector database for semantic search, rewriting how acquisitions are underwritten in the Southwest U.S.

• →Autonomous multi-source ingestion
• →Semantic retrieval over vector store
• →Deterministic site scoring & deal memo

Harbor — 49 Skills, One Router, Flat Context

Harbor provisions skills and MCPs by file tree rather than by agent. An agent resolves only the skills inside its own room at call time, so context cost stays O(1) no matter how large the library gets. A single MCP router with 8 tools — Claude Code, Cursor, Codex, Gemini, Goose — enforces room scope, token budget and audit from one gate. Auto-generated beacons (AGENTS.md, CLAUDE.md, .cursorrules) keep every agent reading the same routing map.

• →49 production skills across 12 categories
• →Skills scoped by directory, not by agent
• →One MCP router — six clients, one gate
• →Self-syncing beacons: every agent reads one map
• →New skills ship without re-architecting agents

The Self-Starting Orchestrator

A meta-agent that owns a build rather than reporting on one. It holds meta-skills — it provisions its own tooling, starts work unprompted, and re-decides its workflow as new events arrive. Nothing waits on a human to type: schedules, webhooks and deltas fire it, a proactive scanner watches the vault and connected accounts for work worth doing before anything has gone wrong, and a canary trips when a job misses twice its expected cadence — low-risk repairs it re-runs itself. The discipline that makes that safe is a governor: every self-directed action is scored and has to clear a confidence bar, and anything touching money or sending on your behalf is hard-gated until you approve it. LLM proposes, deterministic disposes.

The part that compounds is learning how to learn. Three meta-skills ship on first boot. One governs acquisition: check the pool first and reuse rather than rebuild, decide learn-versus-do, sequence the gap that blocks the most work, stop researching once it can act, and archive skills nothing uses. The second learns a genuinely new subject the way a top performer would — trades journals for the edge cases practitioners actually hit, expert interviews for mechanism rather than summary, definitive books for the foundation, docs and case studies for currency — then distills the durable heuristics, the anti-patterns and the numeric thresholds into an installable skill with a measurable output. The third learns from its own lived experience: a task that fails twice gets diagnosed to root cause, fixed under a bounded retry budget, tested against a sibling input so the fix generalizes, and every dead end banked as a don't-rule. Failure is tuition, paid once. This is the holy grail of agentic builds.

• →Cron · webhook · delta · scanner · canary — never a prompt
• →Transport-agnostic by construction
• →Studies the top 1% of a field, then its own failures
• →No LLM in the send path — ever

TDH AI Labs — Autonomous Operations

Companies bleed six figures a year on staff copying numbers between spreadsheets. But handing that to an unconstrained LLM is financial suicide—pure AI hallucinates math and leaks records. TDH AI Labs bridges the gap: hybrid workflows where code enforces exact reconciliation ($0.01 tolerance), regex validation, and atomic database locks before a dollar moves. Workers are compiled from top-1% industry tradecraft—GAAP rules, statutory codes, practitioner heuristics—and every human override feeds closed-loop learning so mistakes never repeat.

VibeHard — Production Software, Not Prototypes

Turning a prompt into a prototype is now a commodity. Tools like Lovable, Base44, and Replit made vibe coding accessible, but they are built for casual prototyping, not production. To make a demo look good in seconds, they skip testing, skip schema migrations, skip concurrency safety, and disable database security. In 2026, audits revealed over 5,000 vibe-coded apps leaking medical records and financial data—including CVE-2025-48757, where Lovable left database Row-Level Security disabled across 170+ live customer sites.

VibeHard replaces casual prototyping with an automated production engineering lifecycle. Between prompt generation and live deployment sits an unskippable battery of 11 automated compiler gates. If an app carries security flaws, broken migrations, or unhandled race conditions, the deploy fails closed. Nothing deploys until it is mathematically proven seaworthy.

The Reviewer Moat: When an architectural edge case or security judgment call exceeds automated resolution, VibeHard doesn't leave you guessing prompts. It packages the exact ±3-line code slice and routes it to an on-demand verified human specialist (Security, Database, Reliability) to inspect and sign off. You get the speed of AI generation with the rigorous sign-off of a senior engineering team—delivering production-hardened software you own 100%, with zero vendor lock-in and zero per-seat subscription rent.