What You'll Do
– Design and ship the ML backbone of Gini AI Workers – routing, tool selection, reasoning, memory, evaluation.
– Build evaluation and feedback loops – offline evals, online A/B, regression harnesses, human-in-the-loop labeling pipelines.
– Optimize cost and latency across the agent stack: prompt engineering, model routing (frontier ↔ small ↔ fine-tuned), caching, speculative decoding, distillation.
– Fine-tune and/or RAG-tune models for vertical enterprise tasks (invoice extraction, PO matching, ticket triage, forecasting).
– Own the ML infra – training pipelines, experiment tracking, model registry, deployment, monitoring, drift detection.
– Partner with backend + product to turn research into shipped features on a weekly cadence.
– Design and ship the ML backbone of Gini AI Workers – routing, tool selection, reasoning, memory, evaluation.
– Build evaluation and feedback loops – offline evals, online A/B, regression harnesses, human-in-the-loop labeling pipelines.
– Optimize cost and latency across the agent stack: prompt engineering, model routing (frontier ↔ small ↔ fine-tuned), caching, speculative decoding, distillation.
– Fine-tune and/or RAG-tune models for vertical enterprise tasks (invoice extraction, PO matching, ticket triage, forecasting).
– Own the ML infra – training pipelines, experiment tracking, model registry, deployment, monitoring, drift detection.
– Partner with backend + product to turn research into shipped features on a weekly cadence.
Requirements:
– 4+ years of ML engineering in production (not just research or notebooks).
– Hands-on LLM experience in 2025-2026: agentic systems, tool-use, function-calling, RAG, structured output, eval design.
– Strong Python. Comfortable with PyTorch/JAX and one serving stack (vLLM, TGI, TensorRT-LLM, SageMaker, or similar).
– You've built an eval pipeline that actually caught a regression in prod.
– You read the papers and know which ones to ignore.
Nice to Have
– Experience with MCP, LangGraph, DSPy, or custom agent frameworks.
– Fine-tuning (LoRA/QLoRA, DPO/ORPO, RLAIF) on open-weight models (Llama, Qwen, Mistral, DeepSeek).
– Vector DBs (pgvector, Pinecone, Weaviate, Qdrant), reranking, hybrid retrieval.
– Prior work on multi-agent systems or enterprise copilots.
– 4+ years of ML engineering in production (not just research or notebooks).
– Hands-on LLM experience in 2025-2026: agentic systems, tool-use, function-calling, RAG, structured output, eval design.
– Strong Python. Comfortable with PyTorch/JAX and one serving stack (vLLM, TGI, TensorRT-LLM, SageMaker, or similar).
– You've built an eval pipeline that actually caught a regression in prod.
– You read the papers and know which ones to ignore.
Nice to Have
– Experience with MCP, LangGraph, DSPy, or custom agent frameworks.
– Fine-tuning (LoRA/QLoRA, DPO/ORPO, RLAIF) on open-weight models (Llama, Qwen, Mistral, DeepSeek).
– Vector DBs (pgvector, Pinecone, Weaviate, Qdrant), reranking, hybrid retrieval.
– Prior work on multi-agent systems or enterprise copilots.
This position is open to all candidates.
