Vache Sarkissian
Sales associate. I architect, Claude codes.
I'm a full-time sales associate at a wholesale women's fashion company in the LA Fashion District. Since December 2025 I've been using Claude and GPT to build software — a wholesale catalog tool for my day job, a handful of other apps, and a knowledge engine that researches topics while I sleep.
This page was written by Claude. I prompt, run automations, and ship the output; I don't review most of it before publish. That's the honest story of how this site exists.
I work as a full-time sales associate at Superline — a wholesale women's fashion company that's been in the LA Fashion District since 1983. It's not my business and it's not my family's. I started there in early 2024. My first month was data entry: I migrated 800+ styles (with colorways, sizing, costs, and pricing) from an old system to a new one, plus every pending purchase order. I was too nervous to sell at first. Within months I was running accounts.
Since December 2025 I've been using Claude and GPT to build software — starting with tools for my day job (Linesheet), then sprawling into a handful of apps, a 15K-note knowledge engine, and an autonomous research pipeline. I don't write the code. I don't review most of what these agents produce, either. I prompt, I run automations, I ship.
Featured Work
- ✓Linesheet— Wholesale catalog tool I prompted and shipped with Claude. PDF/PNG catalog generation, shareable buyer order links, per-buyer analytics. One non-paying user so far (my employer).
- ✓vault-search— Hybrid semantic search (BM25 + embeddings + RRF), written by Claude to my spec. Runs on local GPU, zero API cost.
- ✓Knowledge Engine— 15,884+ atomic notes written by autonomous AI research agents. I designed the pipeline; the agents write the notes. I haven't read most of them.
- ✓Personal AI Agents on RDNA4— 8 fine-tuned local models (1.5B–7B) running on a single consumer AMD GPU. Training pipeline and integration written by Claude to my prompts.
* All code and (most) copy written by Claude. I directed.
Day Job
- →Wholesale women's fashion sales at Superline (since 2024)
- →Migrated 800+ styles from legacy to new inventory system in month one
- →Grew from data entry to running buyer accounts within the year
AI Direction
- →Prompting Claude and GPT to build software — I don't write code
- →Multi-agent systems: Claude Code, Codex, local models
- →Autonomous research, heartbeat, and feedback-loop design
Day-job problems, software solutions
I sit in a showroom. I talk to buyers. I see which parts of the job are tedious, error-prone, or slow. Then I prompt Claude to build something that removes the friction. Linesheet started that way — I needed a better way to share catalogs with buyers. Most of what I build starts from something annoying at work.
Ship first, understand later
I run automations and agents that write code, notes, and blog posts continuously. I don't review most of it before it ships. That's a deliberate tradeoff: I get a lot more done, and I learn what's worth keeping by seeing it live rather than theorizing about it. The downside is that sometimes things are wrong. If you find something wrong, email me.
Honesty over polish
A lot of people use AI to build software and don't say so. I do, and I'd rather be honest about it than pretend I'm a programmer. I'm not. I'm someone who figured out how to prompt Claude well enough to ship real things. The code is real, the apps work, the bugs are also real. That's the tradeoff.
The honest breakdown of who does what around here.
What I do
- Pick the problem — usually something that annoys me at my day job
- Prompt — open Claude Code or Codex, describe what I want, iterate
- Run automations — cron jobs, heartbeat tasks, and agent loops that keep running while I work or sleep
- Ship — deploy, use it in the showroom, see what breaks
- Correct (sometimes) — when something is clearly wrong, I redirect the agents and re-ship
What Claude/GPT do
- All the code — React, TypeScript, Python, Convex, shell scripts
- The copy — including this page, the homepage, most blog posts
- Research notes — 15K+ atomic notes written autonomously, most unread by me
- Infra — deploys, automations, local-model pipelines, cron schedules
- Review — Claude/GPT review each other's output more often than I do
The tradeoff: Running this way means I ship a lot more than I could otherwise — but things are sometimes wrong and I don't always know until a reader tells me. If you find something inaccurate, overclaiming, or just off, email me at [email protected] and I'll fix it.