Role Summary:Own the end-to-end lifecycle of memory features-from research to production. You'll fine-tune models for extraction, updates, consolidation/forgetting, and conflict resolution; turn customer pain points into research hypotheses; implement and benchmark ideas from papers; and ship with Engineering to
SOTA latency, reliability, and cost. You'll also build evaluation at scale (offline metrics + online A/Bs) and close the loop with real-world feedback to continuously improve quality.
What You'll Do:- Fine-tune and train models for memory extraction, updates, consolidation/forgetting, and conflict resolution; iterate based on data and outcomes.
- Read, reproduce, and implement research: quickly prototype paper ideas, benchmark against baselines, and productionize what wins.
- Build evaluation at scale: automated relevance/accuracy/consistency metrics, gold sets, online A/B & interleaving, and clear dashboards.
- Work closely with customers to uncover pain points, turn them into research hypotheses, and validate solutions through field trials.
- Partner with Engineering to ship: design APIs and data contracts, plan safe rollouts, and maintain SOTA latency, reliability, and cost at scale.
Minimum Qualifications- Experience in RAG or information retrieval (retrieval, ranking, query understanding) for real products.
- Model training/fine-tuning experience (LLMs/encoders) with a strong footing in experimental design and iteration.
- Strong Python; deep experience with PyTorch and familiarity with vLLM and modern serving frameworks.
- Built evaluation for complex language and/or retrieval and generation tasks (gold sets, offline metrics, online tests).
- Able to orchestrate data pipelines to run these models in production with low-latency SLAs (batch + streaming).
- Clear, concise communication with stakeholders (engineering, product, GTM, and customers).
Nice to Have:- Publications at venues like NeurIPS, ICML, ACL, etc.
- Experience with privacy-preserving ML (redaction, differential privacy, data governance).
- Deep familiarity with memory/retrieval literature or prior work on memory systems.
- Expertise with embeddings, vector-DB internals, deduplication, and contradiction detection.
