Penumbra
Investigate connected threats while keeping the defending agent within explicit authority boundaries.
Synthetic telemetry · simulated containment
Inspect the evidence briefI build enterprise AI for decisions where money, risk, and accountability matter—connecting financial-services experience with AI architecture, infrastructure economics, and controlled execution.
Based in Austin. Open to in-office or hybrid opportunities within Texas and meaningful business travel.
Secret and agent-authority scanning, security investigations, customer onboarding, commercial research, cloud authority, governed data, infrastructure economics and conversational knowledge.
Investigate connected threats while keeping the defending agent within explicit authority boundaries.
Synthetic telemetry · simulated containment
Inspect the evidence briefBring customer data into a B2B product through reviewed mappings, dry-runs and controlled migration.
Hosted reference demo
Explore the repositoryResearch companies with source-linked evidence and bounded Salesforce access.
Developer-org proof
Explore the repositoryAn agent can become more capable without quietly becoming more powerful.
Verified AWS DEV lifecycle
Explore the repositoryTurn raw shipment data into traceable Gold data products with explicit quality gates.
Databricks reference run
Explore the repositoryConnect cloud spend, GPU usage and AI cost assumptions to executive decisions.
Modeled FinOps reference
Explore the repositoryConnect maintained business knowledge to retrieval, conversation and voice through managed services.
Low-code architecture case
Explore the repositoryFind leaked keys and risky agent configurations, then keep remediation under human review.
Published GitHub Action · offline Mac/Linux CLI
Inspect the evidence briefView Marketplace ActionFirstSync is a hosted reference implementation with synthetic orgs and precomputed model proposals. Inspect each repository for its validation evidence and operating scope.
Catch credential exposure and risky agent configurations before commit or build.
A leaked key is only
part of the problem.
DeadBolt checks local files, staged changes, git history and container layers for leaked credentials and risky MCP or agent settings. Deterministic rules estimate potential authority and help a human prioritize what to review.
The offline core requires no API key or AI model. Run it from a Mac or Linux terminal, use the pre-commit hook, or add the published GitHub Action to a workflow.
Text, JSON and SARIF reports connect findings to content-hashed remediation proposals. DeadBolt records human decisions in a hash-chained ledger and never uses discovered credentials or executes a fix.
Published v0.1.0 release and Marketplace Action. The synthetic demo produces 30 findings. Sixty automated tests and Linux CI across Python 3.10–3.13 cover detection, redaction, staged scans and approval integrity.
Authority is estimated; effective access remains unknown. Detection has finite rule coverage. Approver names are operator labels, and detecting a completely rewritten ledger requires a separately trusted head. The local AI model is optional; real-model inference remains unverified.
Agents that help defend an organization must themselves be defensible.
No single signal tells the whole story.
The pattern matters.
Penumbra takes its name from the partial shadow: the space between full light and darkness. My fascination with Kepler led me, at Harvard, to write about the music of the spheres. What I carry into security architecture is the importance of looking at the whole system.
The moon has two sides, and so do I: the extrovert who works across executives, engineers, Legal and HR, and the focused builder who turns those conversations into working systems. Penumbra needed both.
In 2017, I was writing Hacking The Human about insider threats, social engineering and cybercrime-as-a-service. Today I build agentic AI. Penumbra connects those threads through evidence-linked investigation, explicit tool boundaries and human-reviewed containment.
Synthetic reference implementation with a keyless deterministic demo and optional Claude backend. Live telemetry connectors and SSO-bound approval remain roadmap work; containment is simulated.
Company research becomes a reviewable decision—not automatic permission to act.
Research the opportunity.
Earn the right to act.
SignalDesk connects OpenAI-powered company discovery, source-linked candidates and a human review queue. Exact-draft approvals, persistent Do Not Contact suppression and explicit research limits keep the workflow bounded.
The Salesforce GET-only preview adds a separate, read-only enterprise integration boundary. The XERJ team selected that implementation for its public code-reference corpus.
In the documented workflow, a person approved an acquisition handoff. Separately triggered research could not verify the contact, so SignalDesk retained Needs evidence instead of producing an approvable message.
In the developer-org Salesforce run, two synthetic Leads produced one blocked and one requiring review. Lead permission metadata reported query access allowed and create/update/delete denied. The preview made no CRM writes and granted no outreach authority.
Local, single-user reference application. Live observations are developer-operated; automated provider tests are mocked. Delivery is simulated. No production deployment or customer ROI is claimed; corpus selection is not certification.
Strategy, governance, and execution—each connected to working code and reviewable evidence.
Make the economics inspectable.
Translate AI architecture into cost, performance, and risk trade-offs that executives can evaluate before committing capital.
Published FinOps figures are synthetic and modeled; career outcomes are separate.
Make accountability visible.
Build explicit authorization boundaries, accountable ownership, and traceable evidence for human review.
Reference workflows and synthetic cases; framework mappings are not certification.
Keep research and execution accountable.
Connect financial and operational context with research tools, supply-chain decision support, and controlled execution.
Historical research and simulated execution; no live trading or investment-return claims.
Model selection is an engineering decision—not a brand preference. Accuracy, latency, cost, context, retrieval, security, and failure behavior determine the fit.
The model is a component.
The architecture is the product.
My work spans agentic workflows, model-assisted engineering, enterprise integrations, and deterministic control layers. I evaluate multiple model ecosystems for specific workloads while keeping policy enforcement, permissions, validation, and human accountability outside the model.
Structured research & software engineering
Bounded AI research, source-linked findings, coding assistance, and validated data workflows.
Inspect a bounded OpenAI workflowComplex reasoning & tool-use design
FirstSync uses Claude for schema mapping. Penumbra offers a Claude tool-use backend behind the same scoped tools and policy gate as its deterministic demo.
Inspect FirstSync and its evaluationsImplementation boundary: The portfolio demonstrates governed agent workflows and bounded integrations. OpenAI and Claude implementations are shown separately; a unified production router is not claimed. Multi-agent orchestration and multi-model routing are distinct capabilities; provider-specific production claims require corresponding code and operational evidence.
“The model may propose. The system determines what is authorized.”
I design for the business consequence—not just the successful model response.
Prior career outcomes; not measured returns from these reference implementations.
Explore the wider portfolio across agent authority, capital markets, supply chain, data engineering, cloud economics and cybersecurity assurance.
Open-source recognition: a bounded SignalDesk Salesforce implementation was selected by the XERJ team for inclusion in a public reference corpus. Corpus selection is not certification or a claim of measured adoption.
AI Solutions Engineering & Technical Sales | Enterprise Discovery to Deployment | Advisory & Governance | Business Value First. Based in Austin. Open to in-office or hybrid opportunities within Texas and meaningful business travel.