Michael Lund.
I build AI tools that clear my own bottlenecks, and I get inside an unfamiliar system fast enough to fix it before anyone asks.
Things I built and ran.
AI Request Intake
People at the firm send informal requests for AI and automation work. I built a full-stack app that turns one of those requests into a structured spec. Claude runs up to three rounds of multiple-choice follow-ups, writes an intake record and a build brief, and routes it through an approve, defer, or reject queue. An approval opens a Jira issue.
How it's built
- Python, FastAPI, server-rendered HTML, and SQLite. Claude structured outputs, with a cached system prompt.
- Entra ID SSO scoped to the firm tenant, signed sessions, and a reviewer allowlist. Deployed behind nginx over HTTPS.
- I built it and piloted it for firm-wide use.
Internal tool. No public demo.
Daily Trading Information
Newer traders used to assemble the morning market-data reports by hand. I automated 17 of those recurring reports. The pipeline pulls from SEC EDGAR, newswires, and exchange calendars and renders HTML drafts before the trading day.
How it's built
- A Python CLI. Each report has its own fetch-and-render step.
- A Claude API layer that tiers from Haiku to Sonnet.
- It runs on a holiday-aware trading calendar through GitHub Actions, with 155 offline tests.
Pilot for desk use. No public demo.
Trillium Support Agent
An internal assistant that answers trader and staff questions from the firm knowledge base, with citations. It abstains when it is unsure and escalates urgent issues over Teams or to on-call.
How it's built
- A Python service. Rule-based P1, P2, and P3 triage runs ahead of the model.
- Retrieval is persona-scoped over a BM25-indexed knowledge base.
- Claude returns structured JSON answers with source citations.
- A golden-set evaluation harness covers 64 tests.
Internal tool. No public demo.
Wage Garnishments: What the Company Must Do to Support Them Correctly
A payroll-tech company gave me this as a take-home during a Chief of Staff interview. I had about four hours and no background in the domain. I read the statutes, mapped what the payroll platform handles against what the company has to own itself, and wrote a recommendation the team could act on.
What the briefing does
- It names the real problem in one line. The math is easy. The hard part is moving money to the right place on the right clock and answering the court.
- It gives a phased plan. Launch the common order type now, run the rest on a written manual playbook, and let real volume decide what to build first.
- Every claim carries a confidence marker and a source. Statutes cite the statute, not a blog.
- It reads for a busy person. There's a glossary, a capability map, a competitor teardown, a set of questions for each stakeholder, and a checklist for the first order.
I built this for an interview exercise. It isn't legal advice, and the company names are blacked out.
Crypto sentiment and random forest strategy
A BTC and ETH strategy that predicts 15-day-average returns and sizes each position by how confident the model is. It runs a random forest on price data, a local FinBERT model on news sentiment, and a Google Trends signal. We backtested it on QuantConnect. It took 2nd in Duke's FinTech Trading Competition. The competition scores on Sharpe ratio, so a high finish rewards steady risk-adjusted returns.
What I owned
- I designed and coded the piecewise importance function. It turns the model's predicted move into a confidence weight, and we blended it with a modified Kelly criterion to size positions. That sizing layer made the final strategy.
- I built the Google Trends signal with pytrends and got it working in the QuantConnect backtest.
- I split the team into role-based workstreams and edited the final report.
The strategy (team)
- A random forest predicts direction and size on 15-day-average returns, which cut the daily noise. We benchmarked it against linear regression.
- A local FinBERT model read news sentiment in-house, with no data sent out. It fed dynamic working-memory features.
- Drawdown and liquidation controls held the backtest positive through the 2021 to 2022 crypto crash, when buy-and-hold dropped hard.
CrowdVolt market-expansion pilot
CrowdVolt has traction in New York and Miami, so the opening is to enter cities where the product is strong but listings are thin. I designed a low-cost experiment to do that: hand out free CrowdVolt-branded reusable earplugs at a major EDM festival in a market like LA, San Francisco, or Chicago. Each case carries a tracked QR code, so a giveaway that solves a real need at the show ties straight to app installs, signups, and transactions.
What I designed
- A market-selection rubric: pick a city with low CrowdVolt listing density, a large EDM base, one concentrated festival, and a full follow-on concert calendar.
- The acquisition funnel and metrics, from cost per scan to customer acquisition cost, new listings, and GMV from acquired users.
- A before, during, and after operating checklist, plus a decision rule to scale, revise, or stop on the readout.
The economics
- Sourced three overseas suppliers and compared quotes on the branded case.
- The recommended vendor runs about $0.18 a unit at 5,000 units for the preferred black molded case, so a real-volume pilot stays cheap.
- Distribution treats promoter and venue permission as part of the test, so a win also builds relationships for entering the next city.
This is the plan, economics, and playbook. It is a proposal, not a campaign I ran.
Trading operating system
Daily P&L is noisy, so I built a system to tell luck apart from skill and work on what I actually control. It started as a daily report card and grew into morning prep, end-of-day reviews, weekly pattern checks, automated reporting, and AI workflows that summarize my notes and flag the mistakes I repeat.
What it does
- A daily report card logs my thesis, execution, risk, mistakes, and state of mind.
- Morning prep and end-of-day reviews run every day, with a weekly pass to catch patterns.
- AI workflows condense my notes and surface the errors I keep making.
- Automated reports track how closely I follow the system and my ratio of green to red days.
Claude market-news feed
I kept losing time reading news during trades, so I built a feed. Web scrapers pull market-moving headlines and language models summarize them faster than I can read the source. Nobody assigned it. I hit the bottleneck and built the fix.
What it does
- Scrapers watch the sources and language models summarize each catalyst as it lands.
- It turns minutes of reading into a few seconds when the tape is moving.
- It set the pattern for how I work now. I see a bottleneck, build the tool, and keep it running.
It's a live trading edge, so I keep the details private. A cleaned-up version will go on GitHub.
Notes aggregator
Trading threw off more notes than I could keep straight. I built a tool that pulls the scattered note-taking into one pipeline so I actually review it. It gave me back about two hours a day.
What it does
- It pulls notes out of a dozen places into one feed.
- It summarizes and orders the day so end-of-day review is quick.
- It freed up about two hours a day I used to spend sorting notes.
Built at work. The code on GitHub is a cleaned-up version.
Sell-through dashboard
I spent a week shadowing a friend's 20-person ticket-resale business to learn how they moved inventory. Then I built them a Claude dashboard that models sell-through targets in the weeks before an event, so they can release inventory over days instead of dumping it. It was the first thing I built for them, and they still run it themselves.
What it does
- It models sell-through targets before an event and guides how to release inventory over time.
- The team runs it on their own. It isn't real-time, and it doesn't place buys or sells.
- It came out of a week of shadowing. Get into the system fast, find the thing that moves it, build the tool.
While I watched a few live drops, I saw the team working off memory and steps slipping. I proposed writing down each person's job per drop. That became the playbook below.
Ticket-drop operating system
I watched a few live ticket drops and saw the team running off memory. Steps slipped under pressure and money got left on the table. I worked with the founder to turn the drop into a repeatable system: seven roles, each with a checklist for before, during, and after, so nothing gets missed when it's fast and live.
What it defines
- Seven roles cover the whole drop: ops manager, research analysts, traders, tech, pricer, inventory manager, and fulfillment.
- Each role has an overview, its responsibilities, and checklists for before, during, and after.
- Ownership and hand-offs are written down. Who decides, who buys, who prices, who confirms.
Difmo, my DJ and events company
When bars reopened after Covid, I started a DJ and events company at Duke. I became the resident DJ at Shooters, Duke's main nightclub, and played events for clubs and organizations across campus. It was the first thing I built and ran as my own business.
Highlights
- Resident DJ at Shooters, Duke's main nightclub.
- One of my mixes played at Coach K's Countdown to Craziness, Duke basketball's sold-out season opener.
- Booked and played campus events end to end, from the booking to the night itself.
Where I've built and operated.
- I'm Head of AI for the firm. I own the AI projects and automations across traders, engineering, executives, and accounting. I report to the engineering lead, and I own the strategy, the priority order, and the delivery.
- I turn discretionary trading workflows into production systems. I capture how traders make decisions, encode that into tools and rules, and own the path from a pilot to firm-wide use.
- I'm the front-facing engineer for the traders. I gather requirements, direct AI coding agents to build, review the generated code, and hand validated prototypes to engineering to productionize.
- I own end-to-end delivery of the firm's AI tools. I design and pilot, ship to production with Entra SSO and nginx where it applies, run the day-to-day AI Ops stand-ups with the engineering lead, and keep Confluence and Jira current.
Trader
- I worked inside live trading systems on a short-horizon U.S. equities book. I watched positions, inventory, and risk limits, and adjusted size and exposure when the book drifted from the plan.
- I traded single-name U.S. equities on a catalyst and event-driven book: earnings, M&A, index rebalance, activist and short reports, biotech, and newsflow. Each position had rules-based exits and defined downside.
- I mentored four junior traders on setups, risk rules, habit formation, and structured review so each could run a consistent process on their own, day to day.
- Built a Claude-powered dashboard for a 20-person ticket-resale business that models sell-through targets in the weeks before an event, giving the team a planning read on how to release inventory over days; the team runs it themselves.
- Collaborated on-site on live buying decisions during launch week, including the Gracie Abrams and The Chicks ticket drops.
- Advise the team on buying across the EDM and dance-music market, applying firsthand domain expertise to flag mispriced inventory and demand trends.
- Owned all product deliverables for a cross-functional intern capstone built on Suncor's business needs (user stories, PRD, and roadmap) and ran daily standups to coordinate the engineering team.
- Managed sprint planning in Jira and Confluence and prototyped in Figma, translating requirements into a prioritized backlog.
- Partnered with engineers to deliver a full-stack digital prototype, iterating on features through continuous stakeholder feedback.
- Crafted a compelling pitch deck facilitating a successful Series B financing round, securing $25MM in investments through targeted market research and product strategy.
- Constructed a 50-slide Confidential Information Memorandum (CIM) deck, offering detailed insights to potential investors.
- Conducted financial analysis, driving strategic decisions to optimize financial performance and growth trajectory.
- Partnered with trading-software company TrendSpider to drive customer acquisition and funded-account growth through an incentivized-funding program.
- Launched and managed targeted digital and social-media campaigns, implementing UTM tracking and analyzing Google Analytics data to evaluate performance and refine targeting.
- Created graphics for tutorial videos and blog posts to support marketing content and user education.
Skills and education.
Skills & Tools
Education
Defined risk, asymmetric payoff.
Figures are realized percentage gains. Past results don't predict future ones.
Let's talk.
I'm looking for an operator seat at an early-stage company: chief of staff, founder's associate, or ops. Email is the fastest way to reach me.