Design for the human, not the model.
The CSM Assignment Engine separates mathematical scoring from the human-in-the-loop workflow — the Bouncer pattern halts automation for high-value Enterprise accounts and routes them to a deliberate manual match.
Methodology
How I Work
A repeatable, governance-first approach to designing and shipping production-grade AI systems.
01
Process
From ambiguity to delivery
A structured, four-phase approach to turning vague requests into reliable, measurable systems.
Discovery & Context Gathering
I start by mapping the operational reality. What are the actual inputs, constraints, and success metrics? I avoid the 'solution-first' trap by spending time understanding the problem space, stakeholder incentives, and legacy system limitations.
Architecture & Prototyping
I translate requirements into scoped architecture. For AI workflows, this means defining the retrieval strategy, agent boundaries, and quality gates upfront. I build lightweight prototypes to validate feasibility before committing to full-scale engineering.
Governance & Guardrails
Production systems require guardrails. I implement strict schema validation, rate limiting, hallucination monitoring, and human-in-the-loop fallbacks. If a system can't be reliably audited, it doesn't ship.
Delivery & Operationalization
I don't just hand off code. I deliver runbooks, monitoring dashboards, and training materials. A successful deployment is measured by user adoption and measurable business impact, not just passing tests.
02
Stack
Pragmatic tooling
I choose boring, proven technology for the core, reserving cutting-edge tools for where they create demonstrable leverage.
AI & LLM Orchestration
Glean Agent BuilderLangChainOpenAIvLLM
Backend & Data
PythonFastAPIPostgreSQLSnowflakeTypeScript
Integration & Automation
SalesforceTotangoZapierCustom Webhooks
DevOps & Observability
DockerVercelGitHub ActionsDatadogSentry
03
Governance
Principles over promises
The non-negotiable rules that keep automated systems safe, accountable, and aligned with business goals.
- Rule 01
Clear ownership: Every automated workflow has a named human owner responsible for its outcomes.
- Rule 02
Measurable success: We define the baseline and target metric before writing a single line of code.
- Rule 03
Graceful degradation: When the LLM API fails or rate-limits, the system falls back to a safe, predictable state.
- Rule 04
Data minimization: We only ingest and process the specific data fields required to achieve the stated outcome.
04
Operating Principles
Judgments I can point to
Four rules I keep returning to. Each one is backed by a shipped system and a linked case study, not just a philosophy.
See the case study
Replace monolithic prompts with state-managed chains.
The Transition Readiness Agent pivoted from a single LLM call to a multi-step, quality-gated agent chain. Latency rose 200 ms; grounding accuracy rose 40% and hallucination dropped below 2%.
Practical execution over academic perfection.
The Assignment Engine's constrained greedy heuristic trades the exact-solver guarantee for dynamic capacity recalculation that runs within Google Apps Script's execution budget.
If it cannot be audited, it does not ship.
Both flagship systems log assignments, scoring inputs, and faithfulness checks — the portfolio only publishes what can survive that kind of inspection.
$ Receipts
Published engineering evidence
The portfolio should survive inspection, not just look polished. These are measured production results, not estimated scores.
Security headers
Asecurityheaders.com grade
Production currently scores an honest A. The scan is capped there, even with CSP warnings for 'unsafe-inline' and 'unsafe-eval'.
Lighthouse mobile
| Surface | Perf | A11y | Best Practices | SEO |
|---|---|---|---|---|
| Home | 97 | 100 | 100 | 100 |
| REACH case study | 98 | 100 | 100 | 100 |
Measured against live production on mobile form factor: home and one representative case study both clear the ≥95 target across every category.
Bundle delta after Mermaid pre-render
| Surface | Before | After | Delta | Reduction |
|---|---|---|---|---|
| OpenClaw architecture | 834.1 KiB | 246.3 KiB | -587.8 KiB | -70.5% |
| Transition Readiness reference page | 833.5 KiB | 245.4 KiB | -588.0 KiB | -70.5% |
Those pages used to ship Mermaid runtime code for static diagrams. After switching to build-time SVGs, the diagram-heavy client payload dropped by roughly seventy percent.
06
Proof
Walkthroughs and references
If you want to see how the work is structured, these pages show the real systems, decisions, and delivery patterns behind the portfolio.
Case Study Reference
Transition Readiness Agent — PRD
A concise product brief showing the readiness system's goals, scope, and operating logic.
Case Study Reference
REACH Scoring — System Architecture
A clear view of the data flow, ingestion layers, and scoring logic behind account health decisions.
Live System Page
OpenClaw — Architecture Overview
The production-facing overview for the multi-agent orchestration system, linked from the work index and included in the generated sitemap.