About the role
Key Responsibilities
- Build memory extraction pipelines: turning conversations, documents, and process events into structured facts, entities, relationships, and episodic summaries using LLMs (Claude on Bedrock) and NLP techniques.
- Design and implement memory consolidation: merging duplicate or conflicting facts, progressive summarization of long histories, salience scoring, and decay/forgetting policies.
- Develop context assembly logic: given an agent task and token budget, select, rank, and compress the right memories into the prompt — balancing relevance, recency, and cost.
- Build the Aiva evaluation stack: retrieval-quality benchmarks (precision/recall of remembered facts), memory-drift and contamination tests, hallucination checks, and A/B measurement of downstream agent task success.
- Prototype and productionize embedding strategies: model selection, chunking, hybrid sparse+dense representations, and re-ranking.
- Partner with the Kubogent team to fine-tune and deploy small models for high-volume memory tasks (extraction, classification, summarization) where frontier-model calls are too slow or costly.
- Instrument PII detection and redaction in memory pipelines, working within AI Gateway/ReVAct governance policies.
- Stay current with memory-architecture research (episodic/semantic memory, GraphRAG, agentic memory frameworks) and translate it into shipped capability.
- 4+ years of software engineering with strong Python, including 2+ years building LLM or NLP applications in production.
- Hands-on experience with prompt engineering, structured output extraction, RAG pipelines, and embedding models — beyond demos, with real users and real failure modes.
- Experience building evaluation frameworks for LLM systems: golden datasets, LLM-as-judge patterns, regression suites, and offline/online metric design.
- Solid ML fundamentals: retrieval metrics, ranking, classification, and enough statistics to design and interpret A/B tests.
- Working knowledge of AWS ML stack (Bedrock, SageMaker) or equivalent cloud ML platforms.
- Strong software craftsmanship: testing, versioning of prompts/models/datasets, and reproducible pipelines.
- Experience fine-tuning small/open models (LoRA/QLoRA, distillation) and serving them efficiently.
- Familiarity with agent memory frameworks and research (e.g., GraphRAG, MemGPT-style architectures, reflection/consolidation patterns).
- Knowledge-graph construction experience: NER, relation extraction, entity linking.
- Multilingual NLP experience relevant to Indian enterprise customers (Hindi, Tamil, and other Indic languages).
- Published work, open-source contributions, or strong writing on applied LLM systems.
- Zero-to-one ownership of the platform layer that differentiates Aivar's entire accelerator suite — memory is the moat.
- Founded by four ex-AWS/Amazon executives; backed by Bessemer Venture Partners and Sorin Investments; AWS Preferred Partner building on Anthropic/Claude foundations.
- Your work ships into large regulated enterprises across telecom, BFSI, healthcare, and manufacturing — at production scale, not proofs of concept.
- Small, senior team with direct access to founders and customers; competitive compensation with meaningful ESOPs.