Hector Moran

The Problem

I own 6 rental properties managed by 3 different companies. Every month, 3 PDF statements arrive — each formatted differently, each using different category names — requiring manual reconciliation before anything useful can be done with the data. The real problem: my accountant needed IRS-standard categories and I needed cash-flow visibility. Two users, two completely different outputs, one messy manual process serving neither of them well.

What I Built

A monthly reporting tool that parses rental PDFs through a 3-agent pipeline (Extract → Normalize → Report), maps every expense to one of 15 IRS Schedule E categories, and generates a professional tax-ready PDF — turning a 1-hour manual job into a one-click report ready for my accountant.

Outcomes

• →Under 1 minute per statement — down from 1 hour of manual reconciliation every month
• →6 properties processed across 3 statement formats (HomeRiver, Fox, Bailey-Foristell) with 100% accuracy
• →Zero manual data entry — accountant receives 6 IRS-ready PDFs at year-end, no back-and-forth

Stack

The Problem

Zillow and Redfin are built to surface everything. That's great for the platforms — more alerts means more engagement — but it's the wrong product for a buyer who's already overwhelmed. The signal is buried in the noise. Real estate agents solve this with judgment, but agent judgment isn't available at 7am, isn't personalised to your exact criteria, and isn't scalable.

What I Built

A daily automated digest that scrapes property listings, summarises changes with AI, and delivers a curated briefing directly to Telegram — so I get actionable intelligence in under 60 seconds every morning, without opening a browser.

Outcomes

• →2 real buyers tested with live profiles — daily Telegram digests running with zero manual steps
• →7 scoring criteria across budget, location, basics, must-haves, nice-to-haves, schools, and deal-breaker penalties
• →0–100 transparent score with visible reasoning — buyers see exactly why a home ranked 85 vs 45

Stack

The Problem

Album artwork is user-generated content that ships straight to listeners. No platform can review every cover by hand, and it can't afford to auto-approve the ambiguous ones either. The hard part isn't detection — it's that the two ways of being wrong don't cost the same. A false approval puts harmful content in front of users; a false rejection just adds an image to a review queue. The product question is where you draw that line, and what it costs you to draw it tighter.

What I Built

A moderation portal with a two-stage AI pipeline. A fast vision pass scores every image across six policy categories and routes by confidence — clear cases auto-approve or auto-reject and cost nothing further. Only genuinely ambiguous images reach a second, more expensive agent that reads the policy document, argues both the approval and the rejection case, and rules. Anything still unresolved goes to a human review queue with the full reasoning chain attached, so a reviewer sees why it landed there.

Outcomes

• →A two-tier cost model, measured not estimated — $0.002 per image when the first pass is decisive, $0.007 when the deliberation agent runs. The expensive path only fires where the answer is genuinely uncertain
• →I caught and fixed my own false negative — an illustrated cover scored 10% and auto-approved because the model treated painted artwork more leniently than photographs. After a prompt revision it re-scored at 40% and routed correctly to human review
• →Every decision is auditable and portable — thresholds are configurable per platform's risk tolerance, and each ruling writes a full audit trail and fires a signed webhook

Stack