Equally comfortable in the codebase as on the roadmap.
I own the full software development lifecycle — from requirements and design
through backend development, QA, and production deployment.
I'm a hands-on technical product manager working in B2B SaaS for the cyber-insurance market. I own the full software development lifecycle for a multi-product platform — gathering requirements, writing and assigning the work, running QA and code review, and shipping to production.
Over the past few years I've taken multiple products from 0 to 1 on a modern Python and Vue 3 stack, integrating large language models (OpenAI and Anthropic Claude) and external threat-data sources into shipped features. I lead a small distributed engineering and QA team across several time zones and act as the central coordination hub where engineering, customer success, and sales converge.
What I like most is being technical enough to build it and product-minded enough to know what's worth building — translating fuzzy business needs into a prioritized, well-scoped roadmap, then seeing it through to the release notes.
Designed and built an automated knowledge ecosystem that captures org meetings, logs documentation, and recalls institutional knowledge on demand — drafting user stories and answering questions across departments. Connects Microsoft Teams, Claude, Confluence, and Jira. Saves the team several hours every week.
Took multiple products from concept to production, owning backend development, data ingestion, report and output design, deployment, and DevOps. Shipped 49+ production releases on a steady cadence with maintained changelogs, while adopting AI-assisted development to take on hands-on engineering across the stack and reduce delivery cycle time.
Led a full UI and reporting redesign of the flagship product, owning the work from design and wireframes through implementation and the customer-facing release — improving how security and privacy risk findings are presented to underwriters and policyholders.
A multi-agent pipeline that turns a single topic into a fully produced, published short-form video — research, scripting, voiceover, image generation, animation, assembly, and multi-platform publishing. End-to-end cost lands around $0.38 per video.
Given a topic from a queue, ARP autonomously produces a vertical 9:16 video of roughly 50 seconds — tuned for YouTube Shorts, Reels, and TikTok — and publishes it with SEO-aware titles, descriptions, and hashtags. The intent was a hands-off content channel where my role is strategy and oversight rather than per-video production. End to end, a run lands around $0.38 and 5–8 minutes of wall-clock time.
The system is one orchestrator that owns the lifecycle of a run, plus ten stateless specialist agents that each do exactly one job and return structured output. No agent makes global decisions or calls another agent directly — everything flows through the orchestrator, which stays a "conductor, not creator": it never writes a script or generates art itself. It holds a single source of truth (RunState), calls each agent in sequence, validates every output, retries once on failure, and finalizes the run as published, partial, or failed.
Eleven stages run in order: lock the next queued topic, load the channel's identity (tone, art style, voice persona), gather a few verified facts, write a hook-first 120–180 word script, narrate it with ElevenLabs TTS, break it into three to five visual scenes, generate one 9:16 image per scene under a strict no-text rule, animate each into a short clip, assemble everything in FFmpeg to within a second of the voiceover, publish to each platform over OAuth, and log status, URLs, and cost back to the datastore.
Python orchestration and RunState management, Anthropic Claude for research and scripting, ElevenLabs for text-to-speech, DALL·E 3 for images, FFmpeg for assembly and audio normalization, and the YouTube Data API with OAuth2 for publishing.
Built with AI-assisted development — architecture and product decisions are mine.
An automated stock-screening and analysis platform built on a multi-model agent hierarchy: a Claude orchestrator handling decisions, with cheaper specialist agents running data sync, screening, backtesting, and performance tracking on schedule.
A fully automated swing-trading research system for a small account. It screens a 24-symbol watchlist every trading day, looks for one of three technical entry setups, sizes each position by conviction, and delivers formatted recommendations to my phone over Telegram — backed by automated data syncing, weekly backtest validation, and performance tracking. It runs on OpenClaw, an open-source runtime for local AI agents with shell, filesystem, cron, and sub-agent tooling.
A Claude Sonnet main agent handles the work that needs real reasoning — orchestration, strategy decisions, the morning analysis, and all user-facing communication. Cheaper Gemini Flash sub-agents handle the routine, well-defined jobs on schedule: data syncing, backtest validation, performance tracking, intraday screening, and the nightly review. Routing reasoning to the expensive model and rote work to the cheap one cut operating cost roughly 35% (about $12.50/month versus $19 all-Claude), and giving each sub-agent only the files it needs trims token use further.
The strategy fires on one of three setups — a support bounce, an RSI oversold recovery, or an uptrend pullback — each with mechanically calculated stop, target, and a minimum 1.2:1 reward-to-risk filter. The notable piece is conviction-based position sizing: instead of a fixed bet, risk scales with the RSI reading at entry (4% of the account on deeply oversold setups down to 2% on weaker ones). That single change improved backtested return about 60% over flat 2% sizing, because the most profitable historical trades clustered on the lowest-RSI entries.
A local SQLite store of 44,000+ daily bars makes iteration fast and offline, so I could sweep 36 parameter combinations in a single session rather than hammering the API. The strategy that shipped is the one that survived that testing — and an equal part of the work was rejecting approaches that looked good on paper. A 100-point scoring system lost money; an MACD filter raised win rate but killed 89% of trade volume and dropped profit 82%; trailing stops got chopped out of volatile names. Three simple signals beat all of them.
All agents coordinate through a single Markdown file holding the live strategy parameters, baseline metrics for drift detection, the nightly market outlook, and a timestamped change log. Because state lives in that file rather than a conversation, any agent can be spun up fresh, read it, and act with full context — which is what makes the scheduled sub-agents reliable.
The system is report-only: it recommends, but a human places every order, which has consistently produced better entry prices than the system's own. Sub-agents run under hard guardrails — no strategy changes without main-agent approval, paper trading only, API keys never written to logs, and forced escalation when performance drifts more than 10% from baseline. Across the first five closed trades on real capital it hit its targets for about +$137 (a small sample worth reading as a track record in progress, not a projection).
Built with AI-assisted development — architecture and product decisions are mine.
Own the end-to-end SDLC for a multi-product platform and lead a distributed engineering and QA team. 0→1 delivery across multiple products on a Python / Vue 3 / PostgreSQL / AWS stack with integrated LLMs and external threat-data feeds.
Primary liaison between enterprise product, legal/privacy, and external regulators (CA, UK). Led remediation of 1,000+ critical noncompliant cloud resources during an AWS-to-GCP migration, cutting monthly infrastructure spend by $5,000+ and resolving compliance issues ahead of deadline.
Designed and delivered adaptive curricula for students aged 3–17, managed classes of up to 20, and mentored new teachers.
Full-Stack Software Engineering Certificate — Kenzie Academy (Butler University / SNHU), 2021. B.S., International Business — Ball State University, 2015.
I'm always happy to connect with people building thoughtful software. The fastest way to reach me is LinkedIn — and I'd be glad to share more detail on any of the work above.