About the role

ServaMind is building an LLM Bridge: a trainable layer that connects language models which do not share the same embedding space or vocabulary. This role is focused on model internals - training, evaluation, and systems work around frozen models and adapters - not agents, RAG, or prompt engineering.

You will work on designing, training, and improving that Bridge, including how it lines up with model tokenizers, how it runs across GPUs, and how we measure whether a version is good enough to ship.

What you will do

Design and train Bridge / adapter components between frozen language models (embedding and latent paths, not application glue).
Decide tradeoffs in how the Bridge connects models (for example reconstruction, decoding, or injection of intermediate representations) and back those choices with losses and metrics.
Keep tokenizer and vocabulary identity correct across models: no silent mismatches; version artifacts with Bridge weights and configs. Work with standard Hugging Face tokenizers and with ServaMind's custom encode path where it is needed.
Train and run models on multi-GPU setups (data parallel, tensor parallel, device sharding, FSDP/ZeRO as we use them). Debug training and serving issues with concrete evidence.
Design and run experiments and benchmarks: quality, capability, and systems metrics (accuracy, perplexity, latency, memory, throughput). Set clear go/no-go criteria for Bridge versions.
Optimize training and inference for quality, speed, and memory based on measured results.
Write production Python/PyTorch: clear modules, tests, docs, and versioned artifacts - not notebooks alone.

What this role emphasizes

This role is centered on ML systems and model internals (adapters, embeddings/latents, multi-GPU training, eval). Experience with agents, RAG, chatbots, or prompt engineering is welcome as background, but it is not the main focus of day-to-day work.

If your strongest recent work is application-layer LLM tooling, we still encourage you to apply when you also have depth in representation learning, frozen-model adapters, or multi-GPU training and want to work closer to the model.

Required qualifications

Strong foundation in AI and mathematics: linear algebra, probability/statistics, optimization, and enough representation learning to reason about embeddings, latents, and losses - not only library APIs.
Strong Python and PyTorch for training loops, custom modules, and debugging OOMs / NaNs.
Experience with frozen language models and trainable adapters (or equivalent frozen-backbone + adapter systems), including work on embedding tables or intermediate representations.
Experience with representation learning (autoencoders, discrete or continuous latents, embedding compression, or latent alignment) that you can apply to token embeddings.
Working understanding of how tokenizers and vocabularies affect a frozen language model (encode/decode, special tokens, vocab mismatch) - willingness to work on ServaMind's encode path. Implementing a tokenizer from scratch is not a hard requirement.
Experience training across multiple GPUs: at least one of DDP, FSDP/ZeRO, DeepSpeed, or tensor parallel - with concrete GPU counts and a failure you have fixed.
Experience running experiments and ablations, then optimizing for quality, speed, or memory with clear metrics.
Ability to design software, not only notebooks: modules, tests, versioned artifacts.
Comfort explaining designs on a whiteboard: tensor shapes, losses, and math - without relying on product buzzwords.

Preferred qualifications

Knowledge of or experience with HDC (hyperdimensional computing) / hypervector representations.
Open-source contributions to model training, transformers, inference, or tokenizer tooling.
C++ or CUDA for performance-sensitive paths.
Multi-node LLM pretraining or continued pretraining (beyond single-GPU fine-tuning).
Experience with kernels, graphs, or inference systems at scale.
Papers or production systems on latents, embedding compression, adapters, or vocab adaptation - with a clear statement of what you built.

Artificial Intelligence Engineer

Job description