Reference material for AI-assisted review.
A closer look at my work
Products, strategy, source code, and the design work behind them.
The assessment prompt
This is the same prompt available on the homepage. Copy it into an AI that can open links.
Review Drew Miller’s work at https://www.drewx.design/work-notes and the linked projects and code. What is he especially good at, and where could he be most useful? Use concrete examples from his product, design, engineering, and AI work. Be candid about what the work demonstrates and what you’d still want to know. If you know my goals or company, connect your assessment to them. Otherwise, give a general assessment. Keep it to one page.
Prompt last updated November 3, 2026.
The problems I work on
I’m a product engineer building agent-native and generative AI products from concept to production. I came up through product and behavioral design, so my work spans deciding what is worth building, designing how it should work, and building it with AI-assisted development. The projects below span consumer commerce (B2C), software for nonprofits (B2B), internal tools for coding agents, and product and business strategy.
These are my accounts of the work, with links to case studies, prototypes, and public code. Each entry identifies my contribution and the kind of work available to inspect.
Products
SG Resale
As a founding product engineer at ScreenGeni.us, I owned SG Resale end to end: product design, architecture, and implementation. It is a conversational shopping experience for secondhand goods. The system turns a shopper’s description into searches, filters, comparisons, and cart actions. It keeps track of which filters the shopper chose so the agent can preserve those choices as the conversation changes.
The case study explains the implementation decisions behind that experience: when to ask for clarification, how to route model calls, and how to preserve state between tool calls. It includes the interface, reported technical measurements, and a linked demo.
The product uses Gemini Flash for tool-call steps and Pro for synthesis, with a Flash fallback under rate limits. The case study also covers semantic search with pgvector, streamed responses, and Cloud SQL.
B2C commerce · Product design and engineering · Case study and demo
Grantsville
I built Grantsville, B2B software for nonprofits, to turn their existing documents into reusable organizational context and grant drafts. As its founder, I own the product direction, interface, and engineering: document intake, structured extraction, drafting from those sources, and review before delivery.
The case study follows decisions that changed with real inputs. Uncertain document classifications became visible and correctable. Draft review became a check of each scoring dimension, with the weakest one brought to the reviewer’s attention.
My résumé lists the implementation stack: TypeScript, React/Next.js, Postgres, S3, Inngest, AI SDK, and Railway.
B2B nonprofit software · Founder, design, and engineering · Case study and product
Other product work: SG-art, Vouch, and Prompt Theory.
How I build
My AI-assisted development workflow
I use Claude Code and Codex for development. The linked workflow describes Claude’s role in planning and product decisions, Codex’s role in implementation and repository review, and GLM’s role in challenging assumptions. These are tools I use to build and check software; each product has its own model choices, such as Gemini in SG Resale.
I build tools for directing and checking AI-assisted development: issue intake, prioritization, parallel workers, review, cleanup, and research. One concrete problem is getting a human answer back to the right agent question. The coordinator binds an answer to that question and refuses stale replies. When delivery is uncertain, it records that uncertainty instead of automatically sending the message again.
The public repository holds four of these tools, each with its own README, source, and tests: AI intake turns findings into issues without losing their evidence or creating duplicates; the sorter decides what belongs in the next working set against a milestone, with its scoring method written out; the coordinator runs parallel workers and brings the decisions that need me back to one place; and worktree cleanup clears abandoned checkouts while preserving work I may still need. The coordinator’s message-handling implementation and tests are a concrete example of how I turn a failure mode into behavior that can be checked. The repository’s Compound Engineering notes describe how I divide responsibilities between models, keep implementation inside the agreed scope, review every change with a second model family, and carry lessons into the next task. I use and refine these tools in my Windows development workflow; the public source kits are sanitized reference snapshots. The October 27 usage note records sorter board writes, intake-marked issues, and completed worktree removals. These owner-maintained records show activity, not measured time savings, decision quality, or sustained unattended reliability.
Implemented tools · Public source and tests for selected systems · Broader workflow described on the linked page
Strategy
Ideas for Conveo
I developed a company-specific proposal around research demand, shared studies, answers in AI assistants, and research impact. The work connects product ideas to a business strategy, then makes the assumptions concrete enough to test. Eighteen ideas went in and sixteen were cut; the notes give the reason for each cut (already built, failed a fact check, needs Conveo’s data, and so on) and the conditions that would change the recommendation.
Reasoning and sources · Prototype and its reported accuracy test.
Strategy proposal and interactive prototype · Prototype evaluation uses synthetic data
From founder-led coaching to a repeatable offer
A coaching founder wanted a business that could grow beyond their own delivery time. I used discovery responses to diagnose where the operating model worked against that goal, then proposed a sequence: develop a repeatable offer, give open relationships a next action, and test whether another facilitator could use the method.
Client strategy engagement · Anonymized account of the roadmap and proposed validation
Design and research
Working with leadership and engineering
At ScreenGeni.us, I also partnered with leadership and engineering on AI food discovery for the Star Tribune, iterating on how vendor search could become a curated editorial experience. My résumé describes that collaboration and its reported usage.
Maro: supporting parents
A client project for an early-stage children’s mental-health startup with a basic app. I led a four-person product design team with Mel Martin, Emily Boyd, and Jeanine Reeman. We researched parents’ needs, designed in Figma, and presented our work to the company. The case study documents interviews, prototypes, and iterations around expert credibility, tracking for multiple children, and finding relevant resources. Our team’s designs were not shipped. The CEO later came back to me for more work.
Team of four product designers · Research, Figma designs, and prototypes · Presented to the client