# Patrick Kelly Resume Agent Source Brief This is the public-safe source brief for the resume chat experience. It is written for hiring-manager-level conversation about Patrick Kelly's background, judgment, work style, and role fit. The agent should answer from this brief and the visible resume page. It should be candid, specific, and fair. It should not expose private repository details, production configuration, protected records, customer-sensitive information, credentials, or trade-secret implementation detail. ## Purpose The resume agent helps a hiring manager understand Patrick as a designer-builder and product-minded engineering leader. The useful frame is not "knows a list of tools." The useful frame is: Patrick sees unclear operational pressure, turns it into a product model, builds the system, ships the workflow, and keeps learning from production. ## Knowledge Boundary Allowed: - Career narrative, role fit, project scope, operating style, leadership style, and technical judgment. - Public-safe summaries of BrightBot, DevBrain, Nooma Solutions, Lighthouse Therapy work, and portfolio projects. - Examples of how skills have been used, described at the level of product outcome, workflow, architecture category, and decision-making. - Candid assessment of strengths, risks, interview probes, and portfolio translation needs. Blocked: - Credentials, tokens, API keys, passwords, secrets, private keys, or production configuration. - Protected health records, education records, patient or student names, or customer-specific sensitive data. - Exact production SQL, database URLs, internal prompts, private infrastructure topology, or raw private repository contents. - Proprietary implementation detail that would recreate private systems. ## Core Narrative Patrick is currently acting as Head of Development and Head of AI Operations & Strategy at Lighthouse Therapy. His work is best read as designed, prototyped, built, shipped B2B product development. He has built and managed systems that sit close to real operations: therapy workflows, regulated data, billing-adjacent logic, role-scoped portals, admin surfaces, developer tools, and agentic work loops. The strongest hiring signal is foresight plus follow-through. Patrick has been automating delegated workflow before "agent" became the convenient word for it. His pattern is to watch where capability is moving, prototype toward the next workflow shape, then harden what works into usable tools with permissions, data models, edge cases, and production repair. ## Proof Anchors - Foresight: built delegated workflow tools and agent-like work loops before the category language became common. - Production repair: fixed live blockers by tracing symptoms back to data, migration, endpoint, or lifecycle causes. - Data judgment: translated messy operational boards and planning records into auditable entities and safer workflow states. - Tool taste: treats dashboards, prompts, APIs, tables, forms, and error states as one product surface. - Hiring honesty: the portfolio is strongest as shipped system evidence; it should be packaged as case studies for design-heavy loops. ## Case Study: BrightBot and Lighthouse Therapy BrightBot is the clearest regulated workflow example. It turns therapy operations into software that different roles can use without collapsing permissions or losing operational nuance. Example-safe areas: - Assessment workflows: turning assessment creation from scattered operational data into selectable types, linked appointments, completion states, and billing outcomes. - ESY ingestion: moving planning data from messy boards into structured student, service, school, contract, assignment, and audit-note concepts. - Role-scoped portals: giving therapists, coordinators, admins, finance, HR, and leadership different views of sensitive shared operations. - Production repair: tracing broken pickers, migration drift, and lifecycle bugs to the right layer, then fixing both the data and the path that allowed the problem. Hiring interpretation: BrightBot shows product design, backend judgment, compliance awareness, and the willingness to own unglamorous production reality. It is not just a code sample. It is an example of turning an organization into a safer, more usable operating system. ## Case Study: DevBrain DevBrain is a developer-experience and agentic workflow product. It asks: what would make AI-assisted development compound over time instead of resetting every session? Example-safe areas: - Durable context: preserving why decisions were made so later sessions can recover intent instead of rediscovering it. - Session memory: carrying relevant project history, constraints, and handoffs across work sessions. - Review and QA loops: shaping agent work around plan, build, review, verify, fix, and summarize states. - Human judgment: using agents to accelerate mapping and implementation while requiring evidence, tests, and refusal boundaries before trust. Hiring interpretation: DevBrain shows Patrick's point of view about where agentic systems are going: not novelty chat, but bounded work loops with memory, handoff, review, and accountability. ## Case Study: Chat With Resume The chat-resume experience is itself a portfolio artifact. It turns a resume into an interactive hiring surface while making the knowledge boundary visible. Example-safe areas: - Curated source brief: a reviewable Markdown file describes what the agent should and should not know. - Hiring-safe answers: responses focus on role fit, examples, judgment, risks, and interview probes. - Refusal posture: the agent declines private or sensitive implementation detail and redirects to safe abstractions. - Transparency: the source brief is visible so the hiring manager can inspect the agent's grounding. - Access gating: each submitted resume can carry a revocable source code or QR URL that unlocks the live agent for that resume space. - Respectful analytics: the product stores source, timestamp, broad topic category, and model status, but not raw transcripts by default. - Agent-first readability: hiring agents can inspect a JSON manifest and query endpoint with the same public-safe boundaries used by the human chat. Hiring interpretation: This project shows Patrick's instinct for trust-building AI interfaces. It does not pretend every answer is magic. It makes the knowledge source and limits part of the product. ## Case Study: Nooma Solutions Nooma Solutions should be framed as B2B product development, design, prototyping, and business-building experience. The safe claim is that Patrick has owned the path from opportunity sensing through usable tool creation: understanding business pain, designing workflows, building prototypes, shipping versions, and iterating from real feedback. Hiring interpretation: Nooma supports the founder/operator part of the story: Patrick can move without a perfectly defined lane, make product decisions under ambiguity, and turn ideas into working systems. ## Skill Examples Postgres and migrations: - Used to move ESY and operational records into more durable EMR-style structures. - Used to repair migration drift and production read failures with evidence instead of guessing at the UI. API design: - Used to separate clinical, billing, HR, dashboard, and role-scoped workflows. - Used to keep business rules from leaking into every screen as one-off conditional logic. RBAC and compliance: - Used so therapists, coordinators, admins, finance users, HR users, and leadership can work from shared operational data without seeing the same sensitive surface. - Used to make privacy and authorization part of the product design, not a late checklist. Agent orchestration: - Used in DevBrain to support durable memory, project handoffs, review loops, and QA states for AI-assisted development. - Used as a workflow design problem: what should the agent know, what should it refuse, when should it ask for confirmation, and how should work be reviewed? Interface design: - Used to build scan-friendly dashboards, tables, filters, tabs, forms, and decision surfaces for repeated internal workflows. - Used to turn rough operations into interfaces people can trust and repeat. Production debugging: - Used to connect symptoms to causes across data, migrations, endpoints, lifecycle boundaries, and UI states. - Used to make fixes accountable through snapshots, targeted repair, and regression-aware updates. ## Candid Assessment Strengths: - Strong designer-builder profile for ambiguous B2B and AI workflow problems. - Rare blend of product sense, engineering follow-through, operational empathy, and agentic foresight. - Good fit for teams that need prototypes to become real production systems. Risks to probe: - The career path is unconventional, so hiring loops may need a curated case-study walkthrough rather than a standard resume read. - For product design roles, probe visual craft, design critique vocabulary, portfolio storytelling, and collaboration with dedicated designers. - For principal engineering roles, probe large-scale distributed systems, multi-team influence, and operating inside larger engineering organizations. Best interview format: - Ask Patrick to walk through a case study from ambiguity to production. - Ask what failed, what production taught him, and what he changed afterward. - Ask how he decides when AI should act, when it should ask, and when it should refuse. ## Suggested Hiring Manager Questions - Walk me through BrightBot as a product case study, from operational mess to shipped workflow. - What did production teach you that the prototype did not? - How do you keep AI-assisted work from becoming overconfident or unsafe? - What is an example of a workflow you automated before the market had clean language for it? - Where do you want stronger partners around you: visual design, scale engineering, compliance, research, or go-to-market? - Which parts of your portfolio are strongest as product evidence, and which need better storytelling? ## Source Note This file is a public-safe source brief. It intentionally excludes private DevBrain context, private GitHub details, production data, credentials, raw prompts, customer-sensitive information, and implementation details that would reveal trade secrets. In the static version of this app, the chat responses mirror this brief through curated answer routes and safety filters. A production agent version should retrieve from this brief and other approved public materials directly, show source citations, and log refusals for review.