Backend + Applied ML Engineer (Guardrails)

About us: We build AI Adoption Platform with 3 AI Products— Guardian AI, Shadow AI (employee AI-tool visibility), LLM Observability, and an LLM FinOps Optimizer — for enterprises that need to see and control how AI is used inside their walls.

The role: You'll own the backend and the models behind our Guardrails engine — the system that detects and acts on policy violations (PII/PCI, ethics/bias, prompt safety) in AI traffic. You'll train and ship small, fast, cost-efficient language models that run as services, and you'll extend our guardrail policy engine to cover Shadow AI use cases — applying guardrail policy to the AI usage our Shadow AI product surfaces.

What you'll own

• Training, fine-tuning, evaluating, and shipping small language models / classifiers for guardrails (current stack includes Llama-class detectors, GLiNER + spaCy for PII/PCI, and lightweight safety detectors).
• The inference services that serve them: containerized model microservices on Kubernetes (Graviton/ARM64), with attention to cold-start, model caching, latency, and cost.
• Building eval pipelines and datasets so model quality is measured, not vibed.
• Extending the guardrail

policy engine — moving from detection to enforceable, configurable policy, including new policies tailored to Shadow AI usage patterns.

• The backend plumbing connecting guardrails to the rest of the platform (worker services, queues, dual-DB writes).

You should have

• 3+ years backend engineering, strong in a typed language (TypeScript/Node, Go, or Python services in production).
• Hands-on ML: you've fine-tuned and deployed transformer/SLM models — PyTorch/HF Transformers, tokenization, quantization, eval design. Not just notebooks; production inference.
• Solid cloud + containers: AWS, Docker, Kubernetes; you can reason about GPU/CPU inference tradeoffs and serving cost.
• Pragmatism about model size — you reach for the smallest model that meets the bar, and you can prove it meets the bar.

Nice to have

• NER / PII detection, content-safety, or prompt-injection/guardrail work.
• ClickHouse or other high-volume analytics datastores.
• Experience with on-CPU inference economics (you'll care about per-event cost).