← James Penz
Applied AI Engineering

I ship production AI and it's still running.

The forward-deployed / applied-AI-engineer profile, at operator level: I embed, build, and ship agentic systems with autonomous decision boundaries, eval suites, observability, and human-in-the-loop governance. Not pilots. Not decks.

LangChain / LangGraphClaude Agent SDKEvals & observabilityRAG · MCP
The bar in 2026

The exact skills the market now calls non-negotiable — shipped, in production.

Forward-deployed and applied-AI engineering is the role every AI company is hiring for, and the bar has moved to agentic orchestration, evaluation frameworks, and observability/guardrails on top of RAG. I didn't add those to a résumé — I built a system on them that runs my businesses unattended.

Forward-deployed-engineer postings rose ~729% year over year; agentic-AI roles ~280%. Eval suites that catch hallucinations and regressions before production are now non-negotiable.

— Stanford 2026 AI Index; FDE market reports, 2026
Flagship

AgentForge — a production multi-agent operating system.

Read it the way an engineering panel would. Built to run four business units autonomously, on the modern agentic stack.

Stack

LangChain, LangGraph, and the Claude Agent SDK. 44+ integrated tools, 6 specialized agent teams, a 9-tier cost-optimized model router across Claude, GPT-4, Gemini, Groq, and MiniMax. 26 production APIs.

Middleware

9-layer stack — retry, prompt caching at 90% cost reduction, context summarization, circuit-breaker fallback, cost tracking.

Evals

Nothing reaches production unscored. Hard QA gates block bad output pre-ship; the test agent reads the PRD independently of the code to prevent coverage theater. Full tracing on every run.

Memory / RAG

Dual persistence — Mem0 semantic memory + FalkorDB/Graphiti temporal knowledge graph; hybrid RAG (Voyage AI embeddings + Docling parsing + cross-encoder retrieval).

Fig.1 — AgentForge runtime
Intake6 Agent Teams9-Tier Router 44+ ToolsEvals · QA gates · tracing · 3-tier autonomy governanceMemory + RAG Telegram · HTTPresearch · sourcingmodeling · outreach circuit breakersfallback chainsClaude · GPT-4 · GeminiGroq · MiniMaxQdrant · Mem0 · Graphiti
Breadth

Not one system. Several, in production.

On the implementation-credibility spectrum recruiters use — from "I advised on AI strategy" to "systems still running in production" — this is the top rung.

01

SightForge

Next.js 14 · React · Supabase · n8n

AI sales-intelligence SaaS for a $2B+ industrial manufacturer (70+ facilities, 16 divisions): three-tier AI autonomy, continuous market-intelligence sweeps — 358 qualified targets, 5 new market segments, 60% prospecting-time reduction.

02

Autonomous software-dev pipeline

LangGraph StateGraph

5-agent pipeline — Product → Architect → Code → Test → Review/Heal — with structured artifact cascade and self-healing review loops (~$0.50–2.00 per feature).

03

Financial-modeling engine

Python · native Excel formulas

IC-grade models: DCF, three-statement linkage, sensitivity grids, DSCR validation — formula-driven, not hardcoded.

04

Claude Code skills library

19 versioned skills

Tax prep, financial modeling, premium site builds, 3D rendering, acquisition research — one-off agent workflows turned into versioned products.

The ask

Fifteen minutes. I'll screen-share the architecture and the eval gates, live.