project coco 01: genesis

intro

spending several months in the fashion tech industry has revealed to me (and my team at Inseam) that the industry is riddled with operational problems. the problems we have noticed stretch far beyond fashion, into cpg, and all of manufacturing more broadly. the processes companies at different scales use remain fairly manual, arduous, and painfully slow.

this is the introduction of Project Coco (originally named after Coco Chanel, but now seemingly taking shape of the coconut emoji). this experiment is inspired by Anthropic’s Project Vend. the hypothesis we are trying to prove is:

can Coco take a tech pack, and produce a product end-to-end.

the ultimate goal of this experiment, through our partnership with Singla Intimates, is to develop their swimwear collection and produce a one piece women’s swimsuit. yes, this means getting the package shipped to our doorstep, and (have my cofounders) actually being able to wear it.

background

first some background on what is normally required to create a garment sample.

current manufacturing process for fashion brands.

the creative director or designer of the brand will ideate through mood boards, and come up with an initial concept. this idea for a garment is then translated into a tech pack, which is a document that contains key manufacturing instructions like construction details, grading specs, print placements, etc.

next, reliable manufacturers that align with the desired garment style need to be identified, vetted, and contacted. normally the most reliable way of doing this is by being referred. otherwise, common methods include scouring websites such as Alibaba, 1688.com, ImportYeti, MakersRow, etc.

after some back and forth, confirming vendors for potential candidates, they might send some fabric and colour swatches you might be interested in, in which case you can give the green light, provide payment, and agree to create sample #1 based on the provided tech pack.

usually during this manufacturing process, many brands are ghosted, or manufacturers are inconsistent with their communication. i’ve spoken to brands that have spent over a year to find their manufacturing partner.

after answering clarifying questions and confirming the style you want, the first sample will be shipped, which you can continue to iterate with the manufacturer until you get to your final salesman sample, place a bulk order, and sell to your customers!

technical setup

the idea is to eliminate the majority of the manual process with our very own production agent, Coco.

basic architecture for the Coco experiment.

we (Inseam & Singla) will be communicating with Coco via Slack, the same way we’d communicate with any other team member.

Coco has its own Gmail account, which it will use to communicate with vendors and any external parties. Coco can search through email history, read attachments, and respond professionally to any email.

sneak peek at conversation with tungsten manufacturer during initial testing.

Coco also has access to its own (growing) set of tools such as web search, creating python scripts, general file input/output, etc.

this is the high level architecture, and it is subject to change. as development progresses, i will be adding a more detailed technical breakdown of key decisions and additional tools being added.

future

on one hand, we are trying to see if this will solve a business challenge. from a technical lens this is also an exploration of agent harnessing techniques in a real-world high-stakes environment. this is a deeply underexplored field, and we will see much more developments made in the coming year.

Coco is operational and ready to start working, so we’re going to be putting it to the test: finding manufacturers, negotiating samples, and navigating the messy reality of garment production. you can stay up to date on the process on the website.

i’ll be documenting the journey here: what works, what breaks, and what surprises us. expect follow-up posts diving deeper into the technical architecture.

coco token tracker

to finish off, here is the daily token tracker, Coco was hard at work consuming ~11 million tokens, costing us about $32:

stay up to date on random updates that i post on x!