Prox

A big part of Prox is AI agents that process complex technical documents into structured knowledge. The agents are right most of the time. When they're wrong, we need you to catch it. You'll work inside a review platform we built. Each task shows you the source material, what the agent produced, and the steps it took to get there. You compare them and grade the agent's work. what you'll juggle 1. Read the source and the agent's output side by side. Verify the content was captured accurately. 2. Review what the agent did. What it created, changed, or left out. 3. Score a short rubric covering accuracy, coverage, organization, and rule adherence. Full rubric provided at onboarding. 4. Write detailed feedback about the mistake. This is the most important thing you produce since we use it to improve the agent. 5. Submit. Move to the next task. conditions: - Subject matter shifts over time. You don't need prior knowledge of the subjects. You need to be able to compare two documents carefully and spot where they disagree. - Rate is fixed for the engagement. If it changes, it goes up, and we tell you before your next task. - Work product owned by Prox (work-for-hire). - Standard NDA at offer stage. skills required: - Strong written English - Can read dense technical content for hours without losing focus - Consistent scoring and clear, specific feedback - Reliable on committed hours preferred: - Prior AI trainer/evaluator experience (Outlier, DataAnnotation, xAI, Surge, Mercor, Invisible, Toloka) - Technical writing, editing, QA, translation, paralegal, or research background

Prox builds AI technical support for complex physical products (think power tools, powersports equipment, agriculture, heavy machinery -- anything that requires an installation guide or complex manuals). backed by Y Combinator, Bloomberg Beta, Paul Graham, SV Angel, Burst Capital and many more. intro to the role We run the company out of a git-based knowledge graph. Every customer meeting note. Every investor and vendor conversation. Our writing style, our voice, icp, our positioning. Product ideas, platform new features and roadmap decisions. Marketing campaigns, LinkedIn content, outreach sequences, conference pipeline. Structured, linked, versioned, with an agent layer on top that reads, writes, syncs, and briefs and completes tasks. This is what we build for our customers: take everything a company knows about its products -- manuals, specs, forum threads, support history -- and make it queryable by an AI that never forgets and never sleeps. We do the exact same thing for ourselves. The person we're looking for has a specific character flaw: they cannot do the same thing twice without immediately wanting to eliminate the second time it happens. Not from the laziness, but because repetition feels like a design failure. They think like a hacker. Every system has a seam, and if you find it, you can usually make 10x the impact with a fraction of the effort. They don't ask permission to automate. They notice the pattern and close it. But they have taste. Output that looks like AI slop -- bothers them. Be it an email, customer proposal, blog posts or event announcement. The aesthetic of the work matters a lot as we're entering the Brand Age. Your job is to be the judgment layer on top of everything the agents can't decide -- and to keep extending the system so the list of things agents can't decide gets shorter. what you'll juggle (disclaimer: most of this work exists to eventually eliminate itself except the last two) GTM and BD ops -- pipeline, outreach, follow-ups, proposals executive assistance -- scheduling, travel, logistics, occasional personal tasks content and marketing ops -- LinkedIn, campaigns, contractor coordination video content -- capturing, coordinating, and helping shape what gets made knowledge graph maintenance -- keeping the company ontology clean, current, and growing agent development -- extending the system so fewer things need manual handling skills required: obsessively detail-oriented and caring comfortable in a terminal -- git, basic unix commands, CLI tools + Claude Code comfortable with APIs -- you don't need to be an engineer, but you need to know how systems talk to each other fast learner -- productive in a new system within days strong writer -- natural voice, clean formatting, can quickly catch AI slop, drafts that don't need editing taste -- cares about quality and aesthetic of work systems thinker -- sees a recurring task as a design flaw, not a to-do item Bonus : prompt engineering experience Comp: $135K + 0.35% equity $11,250 signing bonus · full health, dental, vision In-person SF, 6 days/week Process: Application → founder call → take-home challenge → paid work trial → offer

tl;dr – if you're delusional enough, never satisfied with your skill level, insanely hardworking, and want to carry a crazy amount of responsibility, where what you code on Monday impacts thousands of people on Tuesday, this is for you. – you'll work on knowledge engines and multimodal agents for complex physical products, including expressive voice AI and codegen systems that explain what text alone can't, then lead the deployment of all of it to real customers. – no intro calls. to get considered, you post your submission to our challenge. if it's great, you'll hear from us. from there, three more steps. pass them and you're in. – $200k cash salary + 1% equity. – we're backed by Paul Graham, Bloomberg Beta, Burst Capital, SV Angel, Weights & Biases, and more. Position is in person in SF, 6 days/week. ! what we're doing Every product ever sold needs support at some point. That support falls into one of two buckets. Bucket 1: Simple stuff. T-shirts, screen protectors, keyboards. You buy it, it shows up, maybe you ask “where's my order?” once. This is solved. Zendesk and a hundred other horizontal companies solved it. Bucket 2: Hard stuff. $20,000 industrial heaters. HVAC systems. CNC machines. Car parts. Products where buying wrong means your building doesn't have heat or your manufacturing line is down. Support for these products can only be performed by highly trained domain specialists and there aren't enough of them. If you're selling EV charging stations, your support person needs to be a certified electrician who understands local power grids, installation codes, and compatibility matrices. You can't hire this off the street. You can't outsource it overseas. You'd think LLMs would have solved this by now. They haven't. Three years into the LLM era, penetration in this industry is very low & the reason is twofold. First, off-the-shelf models don't actually understand these products. The knowledge lives in 48-page technical manuals buried on some manufacturer's website in terrible formatting — wiring schematics, compatibility matrices, installation diagrams that can only make sense visually. A general-purpose LLM can't draw you the diagram showing how to connect terminal A to terminal B. It doesn't have the spatial understanding or the product-specific reasoning to be a real technical advisor. So companies still rely entirely on human experts. Second, even if the models were good enough, there are no harnesses to make them useful in the business. No engine to capture deep technical knowledge about complex physical products and keep it updated. No way for a company to offload tribal knowledge from their senior technicians into a system. No way to see what questions customers are actually asking and feed that back into the knowledge base. No generative multimodal presentation and no expressive voice support. Prox is building the best technical product expert for extremely complicated physical products. A multimodal agent that can draw wiring schemes, share CAD models, process incoming videos from a technician in the field, and support people over the phone with voice that can pass the Turing test. To get there, we're solving multimodal knowledge graph building at a very deep level. A huge portion of your work will be developing SOTA knowledge engines that can truly understand complex physical products. how we operate – We're always crystal clear with people. You can see it in this job posting — we post the exact numbers you're going to make (no ranges), what it takes to join Prox, and what you'll be working on. It's impossible to build a great company if you're not crystal clear about what you want from your people and where you're going. – In the best spirit of YC, our north star is: are we building something people want? We can't afford to innovate for the sake of it. Solutions we're working on must impact thousands of real people almost immediately. The metrics we optimize are tickets processed, questions resolved, d