Backend Engineer

Join to apply for the Backend Engineer role at Channel3
.

$/yr - $/yr

Channel3 is building a database of every product on the internet. People have wanted to do this for decades, but it wasn’t possible until now – AI is finally smart enough to structure the world’s messy product data, and inexpensive enough to do so  vision: from in-store, to online, to AI-native. We believe agentic commerce is as important as in-store and online channels;

that’s where “Channel3” comes from. We plan to become the central node of agentic commerce, powering every AI transaction and taking a cut of GMV. We see Channel3 becoming as foundational to AI commerce as Stripe is to payments or Plaid is to fintech. The problem:
Agentic commerce is bottlenecked by messy product data. Product data is inconsistent at best, completely wrong at worst — that is, if it exists  opportunity:
McKinsey estimates that “by 2030, the US B2C retail market alone could see up to $1 trillion in orchestrated revenue from agentic commerce, with global projections reaching as high as $3 trillion to $5 trillion.” That future is impossible without great product data.

The team:
Alex (CEO) and George (CTO) have been friends since the first day of Duke. Alex began coding at 9, started his first company at 12, and most recently led AI projects at , where he got this idea. George published research on automating astronomy with robots at 18, and worked on big‑data problems at Palantir for the past 2 years.

Evan and Ignacio are our founding engineers, both joining us from AWS. Evan earned a Masters in CS at Penn and built the first working example of agentic commerce. Ignacio studied CS at Duke with Alex + George and worked on agentic commerce  team is all engineers; we ship fast.

Unfortunately, we are not able to sponsor visas at this time. You must have authorization to work in the US.

Our progress so far:

• We’ve indexed 100M+ products
• 1500+ developers have started using our API
• We’re scoping pilots with some serious enterprise customers
• We’ve partnered with tens of thousands of brands via our affiliate network integrations

• Write code to understand 1B products. We stitch together the latest language, embedding, image, and segmentation models to build a true understanding of each and every product.
  • Product pages vary greatly across retailers. We’ve built a computer‑vision system that can understand any PDP, with no custom code or manual intervention on any site.
  • Deduplicating products across merchants is really hard. Products are described differently, often have different images, and almost never have stable identifiers. We use multimodal models to help us understand different products and match them.
  • Understanding what variants a product comes in might be even harder. We want to know what colors and sizes every shirt comes in, and what configurations customers can order for a new couch.
• Build world‑class search. Developers should be able to find “outdoor grills from Weber, less than $1000, with 4 burners,” “running sneakers under 300g, size 12,” or “this couch [image] but in green” — in under 2 seconds.
  • Embeddings get us 80% of the way there, but we’re always looking to add structured data that lets developers construct powerful queries with deterministic filters.
  • Lately, we’ve been ideating about how to return “good” results, which we’ve learned is not just the highest keyword/semantic match.
• Engineer for reliability and cost. Create evals to measure drift, and guardrails to prevent regressions and hallucinations. Be clever and scrappy to reduce database cost and token usage everywhere we can.
  • We use billions and billions of tokens a month, across self‑hosted open‑source models and flagship models across GCP + Azure. We embed every product, we look at every image with AI, and we run dozens of calls per product — it’s necessary, but it adds up. We’re always looking to reduce token usage, rely on a smaller model, or do some clever context engineering to reduce input tokens.
  • We structure the world’s product data — and we have to store it and search through it (both internally to find matches and to power our API). We’ve taken many steps…