James Penz

The architecture · live, not conceptual

The platform that runs my companies, drawn honestly.

This is not a proposal diagram. Every box below is software running today across four business units, and the live demos on this site run on this exact stack. Start in Business view if you hire; switch to System view if you build. Click anything.

20

agents in production

11

model tiers, 6 providers

4

business units running on it

100%

of model calls traced with cost

Work arriveschat · voice · web · schedulesOne platform20 agents, governed and tracedFour businesses runconsulting · search fund · sites · dealsEverything in the middle box is governed: budgets, heartbeats, evaluation gates, and a human on irreversible actions. Switch to System view for the working parts.

Hover to see connections · click any component for detail · endpoints, schemas, and schedules deliberately omitted

Stack in production

Tools earn their place by running.

No skill bars, no logo wall. Every tool below is grouped the way the platform is layered, with what it does and exactly how I use it. Where it says built, not bought, I wrote it.

Languages & frameworks

The code layer everything else stands on.

Python & FastAPI

35,000-line platform

API framework for the agent platform core.

How I use it: Serves the agent fleet, tool layer, and document analysis as always-on services. Every agent, router call, and pipeline job runs through it.

TypeScript, Next.js & React

5 properties in production

The web layer.

How I use it: Five production sites including this one, the operator cockpit, and the CRM. App Router, server components, print-ready pages, live agent UIs.

Tailwind CSS

Token-driven styling.

How I use it: Each property carries its own design system (this site's Aegean register among them) from shared token conventions, so brands stay distinct on one skill set.

AI & model layer

Six providers behind one router, each doing what it is best at.

Model router (built, not bought)

11 tiers · 6 providers

Routes every task to the cheapest model that clears the quality bar.

How I use it: Eleven tiers covering routing, classification, extraction, reasoning, drafting, research, premium writing, and voice. Tiers swap by configuration; health probes catch degraded providers.

LangChain & LangGraph

checkpointed, durable state

Agent framework and graph runtime.

How I use it: The agent fleet runs as LangGraph graphs with Postgres-backed checkpoints, so long-running work survives restarts. Provider adapters cover all six model vendors, MCP adapters expose external tools, and every call is instrumented into tracing through callbacks.

Anthropic API

runs the demos you can click

Frontier reasoning and writing models.

How I use it: The supervision tier, elite writing, and every live agent on this site: the proposal agent, the readiness diagnostic, the resume tailor.

Google Gemini

Fast multimodal model family.

How I use it: Extraction and parsing on the light tier, multi-step reasoning and research synthesis on the flash tier. The volume workhorses.

Groq

Ultra-low-latency inference.

How I use it: Llama for fast classification where a frontier model would be waste, and Whisper for voice transcription into the task pipeline.

MiniMax

Drafting-class model.

How I use it: Report and code-draft tier, plus latency-sensitive drafting where turnaround beats polish.

Voyage AI

157,000+ companies ranked

Embedding models.

How I use it: Powers semantic memory and the buy-fit ranker that scores the entire acquisition market before a human touches the list.

Data & memory

The system of record and the three kinds of memory agents draw on.

Postgres on Supabase

2 projects · RLS enforced

Managed Postgres, the system of record.

How I use it: Two isolated projects: platform operations in one, the sourcing pipeline and CRM in the other. Row-level security, public keys locked to read-only, writes behind audited API routes.

pgvector semantic memory

Vector search inside Postgres.

How I use it: Past work embedded and retrieved by meaning, so agents recall context across sessions without retraining anything.

FalkorDB & Graphiti

Temporal knowledge graph.

How I use it: Entity and relationship memory: companies, people, deals, and how they connect over time. The who-knows-whom layer under deal work.

Orchestration & automation

What keeps it running unattended, and honest.

n8n

119 playbooks maintained weekly

Workflow orchestration engine.

How I use it: Multi-tenant GTM engines: each client gets an isolated schema, scoped credentials, and workflows stamped from a versioned golden template. Midweek builds, Friday quality sweep, Monday delivery.

Control plane (built, not bought)

20 agents governed

Governance for the agent fleet.

How I use it: Per-agent budgets, liveness heartbeats, and an issue queue with human review. Irreversible actions gate on a person. Autonomy is granted with evidence.

PM2 & schedulers

Process management and cron.

How I use it: Four always-on services plus overnight schedules for research, memory consolidation, and pipeline sweeps. The reason the platform does not need me at a desk.

Playwright

Browser automation for QA.

How I use it: Crawls every page for console errors, overflow, and broken links at desktop and mobile widths; drives user journeys; captured the element screenshots in the gallery on the how-I-build page.

Documents & analysis

Deliverables generated from code, never formatted by hand.

Excel model engine (built, not bought)

1,000+ targets underwritten

Generates institutional financial models.

How I use it: 17-tab, formula-native LBO models: three-statement linkage, native IRR and payment formulas, DCF, seller notes, sensitivity tables. The models an investment committee interrogates.

PPTX renderer (built, not bought)

Generates board-grade decks from a JSON schema.

How I use it: Produced a 68-slide buy-side investment memorandum on the lead acquisition; the client GTM engines produce weekly sales decks with zero analyst hours.

Vision OCR pipeline

On-device OCR for image-only documents.

How I use it: Turned a shelf of scanned engagement decks into a 39,000-word searchable corpus that now governs the copy voice on this site.

Communications & delivery

How work arrives, ships, and reaches people.

Vercel

Hosting and CI for the web layer.

How I use it: Every property deploys through it with preview and production environments, edge network, and per-site env management.

GitHub

Source control and review gates.

How I use it: PR-gated publishing: the weekly SEO article lands as a pull request and publishes only when merged. Nothing auto-publishes.

Telegram Bot API

Chat interface to the platform.

How I use it: Tasks, questions, and approvals from a phone, routed straight into the agent fleet.

Microsoft Graph

Email, files, and calendar APIs.

How I use it: Outbound email from dedicated warmed domains, deal-folder documents, and meeting data feeding the deal-traction models.

Twilio

Programmable voice.

How I use it: The owner-direct dialer in the acquisition engine, prioritizing calls to owners whose numbers survived enrichment.

Cal.com

Scheduling embed.

How I use it: The booking flow on every page of this site, wired to fire the conversion event when a meeting is actually booked.

Shipped with, earlier chapters

Azure ML & cloud services7 AI-enabled processes designed, built, and deployed to production at a $4B manufacturer during the EY years.
Blue Prism & Pega RPAFirst automations into production at a specialty insurer, with process selection and business cases.
AWSCertified Cloud Practitioner; Fortune 500 platform delivery at Capgemini scale.

The method behind the machine: when I use agents, when I use rules, and when I delete the task instead.

How I decide →

How does he decide what to build?

The method