Description
Don't sit in meetings for a mission you don't care about. Build tools to learn faster at
Grasp.
Role
You will be one of the first ML Engineers in the company. You will need to be self-sufficient as, for example, we do not currently have a dedicated MLOps team.
Currently most of our problems are in natural language. We primarily work with self-hosted, fine-tuned, transformer-based encoder models like mpnet, roberta, etc. for problem spaces like classification, semantic equivalence, entity extraction etc. On the generative front, we still primarily use externally hosted models like GPT-4, though we have a business preference to self-host open source LLMs in the future.
You will be responsible for the ML-directed growth and sanitisation of our in-house knowledge graphs. You will also design and iterate on models which leverage knowledge graphs to solve product problems. Specifically, we want you to handle model design, training, and deployment, as an individual contributor.
The work is hard but rewarding. Please apply if you have a desire to get stuck-in building something incredibly useful. Please don't apply if you are looking for an easy ride.
This is a hybrid office position requiring you to be in London regularly.
Company
The mission of Grasp is to increase the rate at which humans learn. A decade from now, Grasp will be able to take anyone from novice to mastery, in any field, ASAP. We will be the centre of learning.
We've raised $4M+ from top tier investors including Balderton, Point9, and Mozilla! Our founders are Ed Matthews and Jacob Sidorov, who built Revolut's trading desk together, and are both avid self-learners.
Grasp is also a member of Makerversity, a pioneering community of over 350 world-leading entrepreneurs, creators and innovators.
Requirements
Essential
- MSc or PhD in Mathematics or Computer Science with at least undergraduate level knowledge of:
- Probability Theory,
- Information Theory,
- Applied Probability & Statistics.
- Mastery of Python and SQL.
- A proven track record of applying traditional and modern NLP approaches to solve real world problems, including several years experience incorporating pre-trained language models into a variety of model architectures and applying further training to them.
- Incredibly familiar with modern encoder-only and encoder-decoder models.
- At least a basic understanding of decoder-only models.
- 2+ years of model training, deployment, pipeline management, maintenance etc. in a production environment.
- Familiar with day-to-day NLP for industry using modern toolchains (SpaCy, HuggingFace, NLTK, etc.)
- Ability to define and obtain required data sets (internally or externally), pre/post-process data, etc.
Nice to Have
You:
- have solved internet search problems.
- have a depth of knowledge in Linguistics.
- have a depth of knowledge in Reinforcement Learning.
- have worked with GNNs.
- have an in-depth understanding of computer architecture, including knowledge of C or C++.
- have worked at a fast-paced startup.
Benefits
- Sign-on stock options bonus.
- A* colleagues with backgrounds at top firms.
- A mission you care about.
- In contact with reality (everything is linked to the user).
