Project Manager - Manufacturing

Sakana AI's aquarium gives enterprise agents a shared brain for business rules. Sakana AI launches Aquarium, an ontology-based agent platform that unifies enterprise data, workflows, and governance for operational AI deployments. · 1 day ago Read 5 min * Sakana AI launched Aquarium, an ontology-based agent platform for enterprise decision support and execution. * Three pillars: Ontology for semantic data, Managed Agents for workflow runtime, Governance for approvals and audit logs. * Workflows are defined as YAML with explicit human-approval gates between agent steps. * Data connects in place from tables, documents, and databases without migration, with live updates. * Built by Sakana FDEs plus veterans from Palantir, Snowflake, Microsoft, and AWS data platforms. * Grew out of deployments with MUFG Bank, SMBC Group, and Daiwa Securities; access is consultation-based. Tokyo-based Sakana AI has introduced Aquarium, an enterprise platform that gives AI agents a shared model of an organization's data, rules, and approval processes. The launch packages three capabilities Sakana previously assembled during individual customer deployments: a semantic data layer, a managed agent runtime, and governance controls. It moves the research lab into commercial infrastructure for building and operating production agents. Three layers share one business model. Aquarium organizes enterprise data and agent workflows around three connected layers: Aquarium's core layers | Layer | Purpose | Key details | | Ontology | Defines business entities, terms, rules, and relationships. Connects tables, documents, and other sources while records remain in their existing stores. Sakana says source updates flow into the shared model so agents can work from current data. | | Managed Agents | Runs workflows that combine retrieval, model judgment, function calls, and human review. Teams describe steps declaratively, and the platform handles workflow deployment and the agent lifecycle. | | Governance | Controls access and records agent activity. Administrators can configure permissions and approval gates. Logs capture decision rationales, actions, results, latency, model usage, and configured accuracy metrics. | The product page represents workflows in YAML, a human-readable configuration format. Each step has a defined role, such as loading data, requesting an agent judgment, calling a function, or pausing for human approval. A single ontology can support multiple workflows, allowing agents to share the same definitions and source connections. That common environment supplies the product's aquarium metaphor. Customer deployments exposed repeated work. Sakana built its reputation through research on evolutionary model merging, self-improving systems, and inference-time scaling. Its AB-MCTS research, for example, explores how additional computation during inference can improve model answers. Aquarium applies the company's work to the operational layer surrounding models. Sakana says the platform grew from deployments in Japanese financial services. Those projects include an AI lending expert developed with Mitsubishi UFJ Bank, automated proposal generation for SMBC Group, and joint development with Daiwa Securities Group. Forward Deployed Engineers working with those customers repeatedly had to connect data sources, reconcile terminology, encode decision criteria, evaluate outputs, and improve accuracy. Aquarium turns that project-specific work into reusable infrastructure. Its engineering team includes people with data platform and governance experience at Palantir, Snowflake, Microsoft, and AWS, a background reflected in the ontology-centered architecture. Semantic gaps become operational risk. Tool-calling frameworks typically give an agent a model, a prompt, and access to functions. Those primitives leave business meaning unresolved. A customer may refer to an account holder in one database, a legal entity in another, and a household in a risk model. Policies may conflict, approval limits may depend on jurisdiction, and identical fields may follow different update schedules. Aquarium's ontology provides canonical definitions for those entities and their relationships. Teams can encode policy precedence, ownership, and approval roles once, then reuse them across agents. This reduces the chance that each workflow reconstructs the organization's vocabulary through prompts and application code. Keeping source records in existing warehouses and document stores reduces duplicated data and migration work. Integration work remains: teams must map schemas, resolve inconsistent identifiers, define permissions, and decide how agents should respond to missing or stale data. Sakana has not yet published enough technical detail to assess how Aquarium handles freshness guarantees, schema changes, conflicting records, or ontology versioning. Explicit approval nodes give regulated organizations a control point before an agent triggers a consequential action. The associated logs can support debugging, evaluation, compliance review, and incident investigation by connecting an output to the data, model call, function execution, and approval that produced it. The closest rivals start from different layers. Palantir Foundry and AIP provide the clearest comparison. Palantir also centers applications on an ontology that connects enterprise data to business objects, actions, and permissions. Aquarium adopts a similar architectural center while drawing its initial customer references from Japanese financial institutions. LangGraph Platform, Databricks agent tooling, and workflow-as-code systems overlap with Aquarium on orchestration, evaluation, deployment, and governance. Aquarium's public materials place greater emphasis on a formal semantic layer shared across many agents. Product maturity, connector coverage, deployment options, pricing, and operational tooling remain difficult to compare because Sakana has disclosed few implementation details. Reusable boundaries for agent systems. * Separate semantics from prompts. Store entity definitions, relationships, and policy rules in a shared layer that multiple agents can reference. * Give workflow steps explicit types. Distinguishing retrieval, model judgment, function execution, and approval makes control flow easier to inspect and test. * Place approval before consequential calls. Human review is most effective when it blocks the action itself and records the reviewer's decision. * Capture an end-to-end audit trail. Logs should connect source data, model usage, tool calls, approvals, latency, and final results. The open questions are concrete. Sakana says it is developing continuous updates for ontologies and agents as organizations change, along with AI-assisted tools for constructing agents. Access currently starts through a consultation form, indicating a sales and implementation process led by engineers working closely with each customer. The public release does not specify pricing, supported connectors, cloud or on-premises deployment options, authentication integrations, APIs, SDKs, rollback behavior, or procedures for testing ontology changes. These details will determine how easily development teams can integrate Aquarium with existing data platforms and delivery pipelines. Sakana has promised case studies in follow-up posts. Evidence for the ontology-centered approach will depend on measurable results such as task accuracy, human override rates, deployment time, latency, operating cost, and policy violations. Comparisons with simpler retrieval-augmented generation systems, which fetch context and then call tools, would clarify when Aquarium's additional semantic layer earns its implementation cost.

Sakana AI's QD-Rec builds smarter shortlists by evolving competing hypotheses. Sakana AI's QD-Rec applies Quality-Diversity evolutionary search to LLM-powered recommendations, growing archives of varied, well-justified picks instead of ranked lists. · 3 hrs ago Read 4 min Subtopic Rag * Sakana AI released QD-Rec, a Quality-Diversity recommender that evolves item plus reason pairs instead of ranking. * An LLM proposer mutates and crosses hypotheses from an archive; a judge scores them under local competition. * A bandit picks candidates, balancing promising ones against unexplored ones, within a tight evaluation budget. * Tested on M&A targets, customer targeting, and paper recommendation, beating fresh-generation baselines on quality and archive size. * Low-scoring hypotheses act as stepping stones; ideas transfer across candidates to produce stronger proposals. * Code is open-source on GitHub under Apache-2.0, runs via uv and OpenRouter. QD-Rec from Sakana AI generates a portfolio of recommendations, pairing each item with a natural-language hypothesis that explains its potential value. The output gives reviewers distinct options and rationales to compare in high-stakes shortlisting tasks such as acquisition research, customer targeting, and literature review. Sakana describes the approach as open-ended recommendation: searching the combined space of candidates and hypotheses under a limited evaluation budget. The GitHub repository uses the Apache-2.0 license, and the research paper is available on arXiv. Scores hide the proposal. A single relevance score cannot capture the different reasons a candidate might be useful. In a mergers-and-acquisitions search, for example, the actionable output is a business scenario connecting an acquisition target to the buyer's products, customers, or strategy. The number of possible scenarios grows quickly because one candidate may support several credible hypotheses. Asking a large language model to enumerate those possibilities across thousands of candidates would require too many model calls. QD-Rec searches selectively, preserving strong ideas while continuing to explore new candidates and rationales. An archive built by local duels. Quality-Diversity algorithms maintain a collection of high-performing solutions across different behavioral niches. QD-Rec applies that model to recommendations, using an LLM to mutate and combine hypotheses while an archive stores the resulting candidate-hypothesis pairs. Each search cycle has three stages: * Select a candidate. A multi-armed bandit balances candidates with promising past scores against candidates that have received little or no evaluation. * Generate a hypothesis. A proposer LLM selects strong archived hypotheses as parents, adapts one to a new candidate, or combines two parents into a new rationale. * Update the archive. A judge scores the generated hypothesis. A hypothesis that fills a new niche enters the archive; one that overlaps an existing niche competes with its nearest neighbor and replaces it only after receiving a higher score. Local competition allows several recommendation strategies to coexist. Even a low-scoring hypothesis can remain useful when it contains an idea that later transfers to a better-suited candidate. What survived 1,000 rounds. Sakana tested the same algorithm across three recommendation domains: | Domain | Candidate pool | Task | | M&A discovery | About 6,300 US companies as of August 2024, represented by short descriptions and seven financial fields | Find acquisition targets for entertainment and financial-services companies | | Customer targeting | 10,000 synthetic California personas averaging about 685 words each | Identify potential customers for products including a 3D printer and an online learning service | | Paper recommendation | About 17,500 accepted papers from recent leading AI conferences | Match papers to a researcher's interests and criteria | The Novelty baseline generated a fresh proposal during every iteration and selected no archived parents. The Novelty+Evo ablation evolved archived hypotheses while choosing candidates randomly, removing the bandit from the process. After 1,000 generation-and-evaluation rounds, QD-Rec produced larger archives with higher average quality in most tested settings. A failed pitch finds a better host. In Sakana's worked example, a vanilla LLM evaluated Liberty TripAdvisor Holdings as a potential acquisition for JPMorgan. It scored the proposal 1 out of 10 and recommended a partnership because Chase already offered a travel product. Elsewhere in the archive, another 1-out-of-10 proposal suggested that Southwest Airlines could embed financing, payments, and loyalty features into the travel-purchase flow. QD-Rec's mutation operator transferred that payments and loyalty concept from the airline proposal to the travel marketplace. The new pairing avoided the regulatory and operational problems identified in the Southwest scenario and scored 5 out of 10. The example shows how a weak archived proposal can supply useful material for a later generation. Run it from the repository. The Apache-licensed repository is a Python project managed with uv. Model requests pass through OpenRouter, using OPENROUTER_API_KEY for the proposer, judge, and embeddings. The default configuration names openai/gpt-6-luna for generation and judging and text-embedding-3-small for embeddings. A minimal customer-targeting run uses the included SkillForge profile: uv sync cp .env.example .env uv run python -m apps.customers.download_personas uv run python -m apps.customers.create_embeddings uv run python -m apps.customers.seed -profile SkillForge uv run python -m apps.customers.qd_rec -profile SkillForge -iterations 100 Paper recommendation requires a Me entry in apps/papers/profiles.py describing the researcher's questions and criteria for a useful paper. The corresponding paper workflow uses that profile name. Runs produce JSON archives that can be inspected in a browser with shared.view_archive. Keep people in the scoring loop. * Candidate context is limited. The current proposer receives short text snippets for each candidate. Richer source data, retrieval pipelines, or a research agent could provide more evidence before generation. * LLM judges have blind spots. Generated proposals may exploit weaknesses in an automated evaluator. Sakana recommends adding human scores for subjective decisions and domains where evaluation errors carry material consequences. Analyst-led workflows can use QD-Rec as an idea-generation and shortlisting layer, with people validating the evidence, assumptions, and final recommendations.