Generative AI on the TensorFlow Stack
Fine-tuning, embeddings, and serving - with honest advice on when to use something else
Generative AI on the TensorFlow Stack
Generative AI changed what a "machine learning project" means, and it changed where TensorFlow fits. We want to be straightforward about both.
DeepAdvisors helps teams build generative and retrieval features on the TensorFlow stack where that stack is the right tool - and we tell you when it is not.
What we do on the TensorFlow stack
Fine-tuning open-weight models with KerasHub. KerasHub (the successor to KerasNLP and KerasCV) gives Keras 3 first-class access to open-weight model families including Google's Gemma. We fine-tune these models on your domain data with parameter-efficient methods such as LoRA, so a single GPU or a modest TPU budget is enough for most tasks: classification, extraction, summarization in a house style, or question answering over your documentation. Because Keras 3 is multi-backend, we can train on JAX for speed and export for TensorFlow serving without rewriting the model.
Embedding models for search and RAG. Retrieval-augmented generation is mostly a retrieval problem. We train and deploy embedding models - fine-tuned sentence encoders or adapted open-weight encoders - that understand your vocabulary, then integrate them with your vector store and ranking pipeline. TensorFlow's serving story is mature here: embedding models are small, stable, and high-throughput, which is exactly where TF Serving shines.
Serving generative models. We deploy fine-tuned models with TF Serving, on Vertex AI, or in your own Kubernetes cluster, with batching, autoscaling, and observability configured. For small models we can also target on-device inference with LiteRT (see Edge AI & LiteRT Deployment).
Evaluation you can trust. Every fine-tune ships with an evaluation set, an automated scoring harness, and a human-review sample. "It seems better" is not a deliverable.
When we will recommend a different stack
- You need a frontier model. If the task requires the capability of the largest hosted models, fine-tuning a small open-weight model will not get you there. We will say so, and we can help you build the retrieval and evaluation layers around a hosted API instead.
- Your team already lives in PyTorch. Keras 3 runs on PyTorch, so there is often a path - but if your training infrastructure, tooling, and hiring are PyTorch-native, forcing TensorFlow into the serving layer adds risk without adding value.
- Training very large models from scratch. That is a JAX or PyTorch distributed-training conversation, not a TensorFlow one, and we will tell you that in the first call.
We would rather lose a project than put you on the wrong stack.
Typical engagements
- Feasibility sprint (2 weeks): data assessment, baseline fine-tune, evaluation harness, go/no-go recommendation.
- Build (4-8 weeks): production fine-tune, embedding and retrieval pipeline, serving deployment, handover.
- Staff augmentation: a senior US-based ML engineer embedded in your team for the duration.
Contact us to talk through your use case. We will tell you honestly whether the TensorFlow stack is the right fit - and what we would do instead if it is not.