About PointFive

PointFive finds the cost buried inside enterprise infrastructure, from the cloud to coding agents. We pioneered DeepWaste across cloud and data platforms, and we are now doing the same for AI token spend with TokenShift.

Fanatics, H&M, Hertz, Nubank, Citizens Bank and other Fortune 500 companies use PointFive to see and cut what their infrastructure and AI actually cost.

We have raised $96 million, most recently a $60 million Series B led by Accel. Engineering sits in Tel Aviv and our go-to-market team is focused on North America. PointFive Labs, our research arm, publishes original work on how infrastructure and AI actually get paid for.

CRN named us one of the 10 hottest cloud computing startups of 2026, Forbes Israel put us on its Next Billion Dollar Startups list, and Redpoint selected us for its InfraRed 100, its ranking of the private companies shaping the future of cloud infrastructure.

About the Role

PointFive is building infrastructure for the next generation of AI-powered engineering.

As AI agents become embedded into developer workflows, the underlying model layer is changing rapidly. Organizations are no longer choosing between a handful of hosted APIs. They are increasingly operating across frontier models, open-weight models, locally deployed models, specialized models, and dynamically routed combinations of them.

We’re looking for an AI researcher to take an integral part in PointFive’s research into how these models behave, how they should be evaluated, and how they can be deployed and used more efficiently across real engineering workloads.

This is a deeply technical research role with direct product impact.

You’ll study frontier and open-source models, develop evaluation frameworks, investigate inference and model optimization techniques, explore local deployment strategies, and help answer questions such as:

Which model should handle a particular task?

How much capability do we lose when we quantize it?

Can a smaller local model replace a frontier API for a specific workload?

How should models be routed across latency, cost, privacy, and quality constraints?

How do we measure whether one model is actually better than another for agentic software engineering?

Your work will directly shape PointFive’s model strategy and the intelligence behind our AI infrastructure products.

What You'll Do

Define and lead PointFive’s LLM research agenda across model evaluation, inference optimization, local models, routing, and model adaptation.

Continuously evaluate frontier and open-weight models across real-world software engineering and agentic workloads.

Design rigorous evaluation frameworks for model quality, reasoning, tool use, code generation, command execution, summarization, and agent behavior.

Build benchmarks that reflect actual developer workflows rather than generic academic tasks.

Research model efficiency techniques including quantization, distillation, speculative decoding, prefix caching, KV-cache optimization, batching, and context management.

Investigate when smaller or locally deployed models can replace expensive frontier models without materially degrading task quality.

Evaluate model architectures, parameter sizes, quantization formats, and runtime configurations across heterogeneous hardware.

Research and benchmark local inference stacks including llama.cpp, MLX, vLLM, SGLang, ONNX Runtime, Ollama, TensorRT-LLM, and emerging runtimes.

Study inference performance across Apple Silicon, NVIDIA GPUs, AMD GPUs, CPUs, Windows workstations, Linux machines, and other endpoint configurations.

Develop model-routing strategies that optimize across quality, latency, cost, privacy, context size, and hardware availability.

Explore intelligent cascades where smaller models handle common tasks and more capable models are invoked only when necessary.

Research model specialization through fine-tuning, LoRA, adapters, distillation, prompt optimization, and other model adaptation techniques.

Investigate model behavior under constrained environments, including offline execution, limited memory, limited compute, and local-only inference.