01The gap
A founder, a student, or an independent inventor can describe what they made in a sentence. Getting that sentence into a patent or a trademark they can actually file is a different problem. The tools around them stop early. A search tells them something exists nearby. A template gives them blanks. A chatbot writes claims that have never met the invention, the drawings, or the classes. Each one is allowed to call itself help. None of them carries the idea the whole way: specification, figures, classes, and a filing a licensed attorney can review.
The loss is the invention that never gets filed, because no product was built to do the whole sequence. Arcangel closes that gap.
The work sits on the application itself. Halo reads what the person built, in plain language or from a GitHub repository, and drafts the filing. Novelty and clearance show what is already out there. A sketch, a photo, a PDF, or a 3D model becomes patent-style line drawings with reference labels. A licensed attorney, independent of Arcangel, reviews where the filing needs a human. Arcangel is not a law firm. The product is what puts a finished application in front of that attorney, and then in front of the USPTO.
The people using it are independent inventors, first-time founders, students, researchers, and the accelerators and companies that back them. A patent or a trademark is how someone with no legal department owns what they made. When the product works, the idea is on file.
02The engineering
This is not a CRUD app with a model call bolted on. You would work across the whole stack, from ingestion to inference to the pixel.
Heterogeneous technical and legal corpora normalized into a canonical schema, with entity and family resolution over linked-document graphs. Hybrid search: sparse BM25 fused with dense bi-encoder embeddings via reciprocal rank fusion, re-ranked by cross-encoders fine-tuned on relevance judgments. LLMs decompose queries into atomic elements, produce element-to-passage alignments with span-level citations, and beam-search over multi-document combinations to build evidence chains.
Unstructured input compiled into a typed knowledge graph of entities, relations and variants, then decoded under grammar-constrained generation against a formal grammar, so every output is parseable by construction and hierarchical outputs are validated as trees. Long-form text generated by retrieval-augmented decoding conditioned on the graph, with provenance tracing every generated element back to its source.
Domain-adapted LLMs through supervised fine-tuning on paired structured and long-form text, then preference alignment via DPO or RLHF on expert redlines, where every human edit becomes a preference pair. Outcome classifiers trained on historical decisions under weak supervision, producing calibrated probabilities that double as reward models for best-of-n sampling and candidate reranking. Embedding-space density estimation to quantify scope and novelty, turning qualitative trade-offs into optimization targets. Per-tenant LoRA adapters for organization-specific style without pooling data.
A symbolic and neural check layer: dependency parsing, coreference resolution and NLI-style entailment to enforce internal consistency, catch undefined references and ambiguity, and confirm every claim in the output is supported by the source.
Multimodal encoders and instance segmentation lift sketches, photos and text into scene graphs of parts and spatial relations. Graph-to-geometry layout emits vector primitives, not raster, so every element is an addressable object under affine transforms, with annotations bound to the underlying graph and rendered to formal conventions.
Multi-modal similarity search: phonetic distance over grapheme-to-phoneme transcriptions, contrastive image embeddings for visual similarity, semantic embeddings for meaning. Perceptual hashing for image and video robust to crops, overlays and re-encodes. MinHash and LSH for near-duplicate text detection. Web-scale crawling with entity resolution feeding thresholded match pipelines with append-only evidentiary logs.
Large public records modeled as both time series and heterogeneous graphs. Citation networks with centrality measures, temporal topic modeling and clustering over document embeddings to detect emerging clusters before they have names, and graph neural networks over entity-citation-category graphs to produce leading indicators. Two-tower retrieval models embedding two sides of a marketplace into a shared space for matching.
Event-driven and event-sourced, with finite state machines modeling long-lived entity lifecycles and rule-based deadline computation across jurisdictions. Horizontally scaled job fan-out, deterministic and reproducible inference, zero-retention model routing, single-tenant deployment inside customer VPCs, and immutable audit logs on every inference call.
You own features end to end: the pipelines and models underneath and the interface a professional relies on when the stakes are real. Concept to deployment, including evaluation, documentation and monitoring. Very few people stand between you and the user, and what you build either holds or it does not.
03Who we are looking for
Builders own the outcome. Workers execute the task. We do not hire workers. You will work harder here than anywhere you have worked before, and you will be measured on outcome, on impact, and on what you shipped. We want someone driven to the edge of obsession. Nobody had to tell them to care. This is the bar.
You are building a company, not working at one. The thing is not done when your part is done. It is done when it works for the customer.
Your output has to be something people can trust without checking it twice. When you say something is done, it is done. Be the person nobody has to double-check.
Say what is true, when it matters, to the person who can act on it. We do not soften things here. The most useful thing you can do for someone is tell them the truth.
If something is broken and it is not technically your responsibility, it is still your problem until it is fixed. We are too small for “not my job,” and we will always be too small for that.
Move before you are asked. Identify the problem before it has a name. The best people here do not wait for a ticket. They see what needs to exist and they build it.
Work hard, compete hard, push hard, and be genuinely happy when the person next to you does something better than you could have. The market is outside. Inside this building we are one thing.
And be present. Be locked in. Chronic lateness, disappearing during the day, coasting, none of it is compatible with this place. Your team is counting on you to be locked in.
We have one standard: did you move the company forward today? If yes, you will be recognized and rewarded. If no, consistently, you will not last here. Every person who is not moving us forward is taking something from the people who are. That is the whole performance review, and it is the same one we hold ourselves to.
04No resume gate
This role is resume-blind. No degree gate. No logo filter. We have never cared about where you worked, and we are not impressed by it. A raw, undiscovered builder who can outbuild a pedigreed engineer is exactly who we are trying to find.
We decide on a hard, real work sample drawn from the actual product, the same one for everyone. You build something real, and we look at how you think, how you decide, how far you push it, and whether what you shipped holds. Show us you can build. That is the interview.
You are fluent with Claude Code, Cursor, Codex, or whatever your tool is. You use it to ship more than one person normally could.
APIs, document generation, and the cloud they run on. The product is React, TypeScript, Vite, and Supabase.
You are the person nobody has to double-check. You finish the last ten percent.
Genuine curiosity about how a filing actually gets made: claims, prior art, figures, or the product an inventor uses. The kind you cannot switch off.
Bonus, not required: a patent or trademark workflow, document generation, or having taken an early product from zero to one. None of it is a gate. The work sample is the gate.
05In person, in Chattanooga
This is an all-in, in-person, multi-year sprint in Chattanooga, Tennessee. Required. No remote, no hybrid. If that is a dealbreaker, this is an easy and honest no for both of us.
The point of being in the room is that the work gets your time. You are with the team, on the product.
We are not going to print an hour count or a schedule, and we would be suspicious of any founder who did. The best builders do not sign up for a clock. They sign up for a mission and a bar, and then they volunteer everything the work requires, because they want it that badly. If you are at a stage of life where vacations matter to you, this is not the right place. If you read long hours, a small team, and high stakes and feel something tighten with excitement, you are the person we are writing to. Arcangel is not a job. It is a training arc.
06Non-negotiables
We currently work out of Brickyard. The role is in person in Chattanooga, with the team, the whole time.
You must already hold eligible work authorization in the United States. This is an in-person role and we are not in a position to sponsor at this stage.
We are intense, and we are honest about it. The bar on this page is how we actually work. If it makes you want to sprint, we should talk. If it makes you hesitate, this is not the right place.
07The deal
Most people spend their careers trading time for income. It ends when you own a piece of something that grows. In exchange for going all in, you grow faster than you would anywhere else, and you hold equity in something with real upside.
You get equity when the company is worth $50 million. We are building this to a $5 billion IPO, and at that point your grant is worth 100 times what it was when you got it. You do not need to win the lottery. You need to be early at the right company and earn your place in it. Illustrative, not a guarantee of value. The offer letter and the grant documents control.
An early founding grant with the upside above. This is the heart of the offer. The offer letter and the grant documents control.
A competitive base, plus the equity grant.
08How to apply
The single best thing you have ever built, and a short paragraph: why it was hard, what only you would have done differently, and why this is the thing you want to spend the next chapter of your life on. No resume. A resume will not help you here.
A hard, real problem drawn from the actual product, the same one for everyone, the kind of thing you would hit in your first month. Paid. We look at how you think, how far you push it, and whether what you shipped holds.
If your sample holds, you come to Chattanooga and build alongside us on the real product for a few days. The only honest way to know whether we should do this together is to actually do it. That trial is where this is decided.