Second part of a two-part series on catching mismatches between product photos and their descriptions in e-commerce listings. Part 1 used CLIP and BLIP as a similarity baseline — this piece adds OCR and visual question answering to catch the mismatches that baseline missed, like wrong colors, sizes, or counts stated in the text.
A LangGraph agent that checks marketplace listings for label mismatches — but instead of running the same fixed checks on every product, it first retrieves the right verification policy for a listing's category, then reasons about what actually needs checking. Picks up directly from a gap named in the earlier image-description alignment project: different categories need different checks, not one fixed pipeline applied to everything.
An allegorical story about a market vendor whose business is wiped out by a new import levy — one sold as protection for local traders but that ends up benefiting a lord instead. A short parable about how policies framed as protecting the many quietly end up concentrating power in the few.
Surfaces smart-money signals from Polymarket — large bets at extreme odds and shifts in crowd consensus — then uses local LLM analysis to map those signals to the stocks they're likely to affect. A side project exploring whether prediction-market activity can serve as an early signal for equity moves.
Adds OCR and visual question answering on top of the CLIP/BLIP baseline to catch mismatches the embeddings alone miss — wrong colors, sizes, or counts stated in the listing text.
Builds a CLIP-based similarity baseline and a BLIP captioning check to flag e-commerce listings where the product photo doesn't match the written description.
The GitHub repo behind the series above — code, notebooks, and data for detecting image/description mismatches in online marketplaces.
A category-aware LangGraph agent for listing verification — it retrieves the right verification policy for a product's category, then reasons about what to check, instead of running one fixed set of checks on every listing.
The closing part — extends the agent from managing one zone to coordinating two at once, using the memory and human-in-the-loop steps built up earlier.
Adds a human-in-the-loop checkpoint so the agent pauses for approval before taking a risky rebalancing action.
Gives the agent memory of its own past actions and their outcomes so it can run unattended for hours instead of reacting one step at a time.
Adds an LLM step that reads free-text operations notes and folds that context into the agent's rebalancing decisions.
Starts the series with a plain rule-based agent for a single ride-share zone — no LLM — to establish a baseline before adding any AI.
An allegorical story of a market vendor displaced by an import levy sold as protection — a parable about who policies framed as protecting the many actually end up protecting.
What one drive across town showed me about my own judgment — and how the status of the person who makes a mistake, not the mistake itself, quietly sets the standard we judge them by.
The future always wins. The question is who profits from it — a satirical look at how dominant companies stay dominant, from Standard Oil to today.
Surfaces smart-money signals from Polymarket prediction markets — large bets at extreme odds and shifts in crowd consensus — then maps them to impacted stocks using local LLM analysis.