Self-hosted copy — sessions-only variant, from local agent-session evidence alone. The combined sessions + GitHub profile is the primary version. Not submitted to Overflow.
David Strawn
Runs small consumer web products by directing coding agents through build, deploy, operations, and marketing, with hands-on output review.
01Project role
The evidence most resembles an operator who translates goals into working systems by directing agents. Across seven projects, the observed pattern is consistent: define the outcome in plain language, let an agent build it, then personally verify the result against the real artifact — listening to edited audio for dropped words, comparing rendered frames against reference screenshots, checking export timing on a production server. Product decisions (disabling a payment flow, deferring a scaling problem, cutting a feature) are made by the person; implementation is made by agents.
It resembles a hands-on full-stack implementer much less — no hand-written code, test authoring, or code-level review appears anywhere in the session history — and it does not resemble an ML engineer or researcher at all; AI use is application-level direction of coding agents, not model work.
02Tools & working environment
claude-codeprimary · 12 sessions
Default agent across all seven observed projects. Delegation spans the full lifecycle: building features, debugging from pasted console logs, deploying, resolving divergent git history, and triaging production incidents on a live server.
browser & computer usefrequent · 4 sessions
Directed in-browser work to reconfigure a Reddit-marketing SaaS and capture its guides; two earlier desktop computer-use sessions delegated shopping-cart automation and explored capability limits. Visual QA via screenshots in a game-UI project.
custom agent skillsfrequent
Custom skills present and exercised in two different agent CLIs: a research-paper corpus toolchain (ingest, browse, full-page fetch) in Claude Code, and a customer-prospecting workflow in Codex. Skill authorship is likely agent-assisted.
codex clioccasional · 1 session
One retained session, invoking the prospecting skill and planning customer outreach. Codex retention may hide earlier use.
git / githubfrequent
Directs commits, pushes, private repo creation, and a keep-ours resolution against a diverged remote. No PR-based or team workflow observed.
production vpsfrequent
Operates a small production server hosting a video product, with a monitoring agent running on it and a managed Postgres instance behind it; directed triage of a 504 export timeout, a database connection drop, and a ~2-hour render that needed to get faster.
cursoroccasional
Present as an editor with many historical workspace folders, but no readable AI-session history; working familiarity is not established from sessions.
03Languages, frameworks & platforms
agent-directed deliveryvery familiar
The consistent mode of work in every session: plain-language task definition, mid-flight correction, and personal acceptance review of the finished artifact. Includes catching agent mistakes the agent missed — duplicate clips, truncated sentences, a suspected visual artifact in rendered video.authorship: direct
agent orchestrationfamiliar
Coordinates multiple agents across products; local transcripts include one session that fanned out a 100+ subagent research workflow, and references to standing agents running elsewhere (a server-side monitoring agent, other product agents).authorship: direct
video / media pipelinesfamiliar
Repeated FFmpeg-centered work across two projects: render-pipeline performance triage, a history of argument tuning for color accuracy, a WebAssembly FFmpeg build in the browser, multi-clip stitching, and transcript-verified editing. Proposed A/B testing pipeline changes on a sliced sample before trusting them.authorship: directed & reviewed
javascript web appssome
Front-end debugging directed from pasted browser console output; the products are agent-built web apps. No hand-written code observed; specific framework depth cannot be established.authorship: directed & reviewed
postgressome
A managed Postgres instance sits behind a production app; directed triage of a connection drop. No schema or query work visible.authorship: directed & reviewed
geospatial datasome
Directed a mapping tool layering zoning status, appraisal values, water sources, and pollution data for regional land evaluation, and reasoned about which public datasets were worth importing.authorship: directed & reviewed
deploys & operationssome
Directs deploys of two small production sites and post-deploy verification. No CI pipeline or infrastructure-as-code authoring observed.authorship: directed & reviewed
04Subject-matter experience
video captioning & post-productionsubstantial
Operates a browser-based video-captioning product (upload, transcription, styled caption rendering, export) and separately directed a multi-clip editing pipeline with transcript-verified cuts and filler-word evaluation. Repeated sessions across two projects.
early-stage marketing & customer discoverysome
Configured a Reddit-marketing SaaS for a product, archived its playbooks locally, and ran a prospecting workflow that identifies and ranks potential first customers, then asked to be walked through outreach. Small number of sessions.
consumer education contentsome
A deployed educational guide on household allergen remediation — restyled, deployed, payment flow deliberately disabled to make it free — plus a small scientific-paper corpus organized around explicit research questions. Two projects, few sessions.
land & real-estate datasome
Regional land-evaluation work combining zoning, appraisal, water, and environmental data for a planned community concept. One project.
05Relative profile
Versus a product-oriented full-stack engineer: covers the same surface area — front end, media pipeline, database, deploys — but through delegation rather than implementation. The distinguishing strength is breadth of ownership (one person carries idea → build → operate → market); the gap is that implementation depth is simply not observable here.
Versus a consultant/operator: this is the closest match. Requirements arrive as plain-language outcomes, tradeoffs get decided quickly and pragmatically, and verification is done against the real user-facing artifact rather than against the code.
Versus an automation/integration specialist: shares the tooling instincts — custom skills in two agent CLIs, browser automation, standing server-side agents — but the automation serves their own products; no client-facing integration or API-design work appears.
Versus an ML engineer/researcher: no overlap in evidence. No training, evals, embeddings, or model infrastructure work; AI usage is entirely at the application-direction level.
06Project match
Good fit
Agent-directed delivery of small web products from idea through deploy
Acceptance review and QA of agent-produced work, especially audio/video output
Operating small deployed services (single-server scale) with agent support
Early-stage marketing and customer-discovery experiments run through agents
Media/FFmpeg pipeline work specified and verified at the output level
Bring a specialist
Performance and scaling beyond a single small server
Front-end or design work requiring hand-crafted implementation
Database and schema architecture
Codebases where a human must review changes line-by-line
Not established here
Direct hand-written implementation in any language
ML model training, evals, or research
Infrastructure at scale (Kubernetes, IaC, multi-service platforms)
Small evidence base. 15 retained sessions total: Claude Code 12 (2026-06-05 → 2026-07-15, one of which is the profile-generation run itself), Codex 1 (2026-07-13), Cowork 2 (2026-04-20 → 04-22). Local retention is rolling; earlier history is gone, and Codex/Cowork volumes almost certainly understate actual use.
One session excluded. A personal session was dropped from all examples for privacy; only tool-usage patterns from it were retained generically.
Cowork fidelity is low. Those two sessions contribute only titles and delegation intent, not implementation detail.
Authorship is almost entirely agent-directed. The session history neither proves nor disproves direct coding ability — it shows the person choosing not to hand-write code while using agents. (The combined profile incorporates GitHub commit history, which does establish earlier direct implementation.)
Cursor shows editor use (many workspace folders) but no readable AI-session data; it contributes nothing to skill claims.
No release proof. Two products are observably deployed and operated on live infrastructure; session logs cannot establish traffic, customers, or revenue.
Cadence (metadata only): 11 sessions in the last 28 days, 7 active weeks of the last 12, most recent session 2026-07-15. Volume is not treated as skill.