The Amazon.com Catalog is the muse of each buyer’s procuring expertise—the definitive supply of product data with attributes that energy search, suggestions, and discovery. When a vendor lists a brand new product, the catalog system should extract structured attributes—dimensions, supplies, compatibility, and technical specs—whereas producing content material equivalent to titles that match how clients search. A title isn’t a easy enumeration like coloration or dimension; it should stability vendor intent, buyer search habits, and discoverability. This complexity, multiplied by tens of millions of each day submissions, makes catalog enrichment a great proving floor for self-learning AI.
On this put up, we reveal how the Amazon Catalog Group constructed a self-learning system that repeatedly improves accuracy whereas lowering prices at scale utilizing Amazon Bedrock.
The problem
In generative AI deployment environments, enhancing mannequin efficiency requires fixed consideration. As a result of fashions course of tens of millions of merchandise, they inevitably encounter edge circumstances, evolving terminology, and domain-specific patterns the place accuracy might degrade. The normal strategy—utilized scientists analyzing failures, updating prompts, testing modifications, and redeploying—works however is resource-intensive and struggles to maintain tempo with real-world quantity and selection. The problem isn’t whether or not we will enhance these programs, however learn how to make enchancment scalable and computerized somewhat than depending on guide intervention. At Amazon Catalog, we confronted this problem head-on. The tradeoffs appeared not possible: massive fashions would ship accuracy however wouldn’t scale effectively to our quantity, whereas smaller fashions struggled with the complicated, ambiguous circumstances the place sellers wanted essentially the most assist.
Resolution overview
Our breakthrough got here from an unconventional experiment. As an alternative of selecting a single mannequin, we deployed a number of smaller fashions to course of the identical merchandise. When these fashions agreed on an attribute extraction, we may belief the outcome. However once they disagreed—whether or not from real ambiguity, lacking context, or one mannequin making an error—we found one thing profound. These disagreements weren’t all the time errors, however they had been nearly all the time indicators of complexity. This led us to design a self-learning system that reimagines how generative AI scales. A number of smaller fashions course of routine circumstances via consensus, invoking bigger fashions solely when disagreements happen. The bigger mannequin is carried out as a supervisor agent with entry to specialised instruments for deeper investigation and evaluation. However the supervisor doesn’t simply resolve disputes; it generates reusable learnings saved in a dynamic data base that helps stop total lessons of future disagreements. We invoke extra highly effective fashions solely when the system detects excessive studying worth at inference time, whereas correcting the output. The result’s a self-learning system the place prices lower and high quality will increase—as a result of the system learns to deal with edge circumstances that beforehand triggered supervisor calls. Error charges fell repeatedly, not via retraining however via collected learnings from resolved disagreements injected into smaller mannequin prompts. The next determine exhibits the structure of this self-learning system.
Within the self-learning structure, product information flows via generator-evaluator staff, with disagreements routed to a supervisor for investigation. Put up-inference, the system additionally captures suggestions alerts from sellers (equivalent to itemizing updates and appeals) and clients (equivalent to returns and destructive evaluations). Learnings from the sources are saved in a hierarchical data base and injected again into employee prompts, making a steady enchancment loop.
The next describes a simplified reference structure that demonstrates how this self-learning sample may be carried out utilizing AWS companies. Whereas our manufacturing system has extra complexity, this instance illustrates the core elements and information flows.
This method may be constructed with Amazon Bedrock, which supplies the important infrastructure for multi-model architectures. The power of Amazon Bedrock to entry various basis fashions allows groups to deploy smaller, environment friendly fashions like Amazon Nova Lite as staff and extra succesful fashions like Anthropic Claude Sonnet as supervisors—optimizing each price and efficiency. For even better price effectivity at scale, groups may deploy open supply small fashions on Amazon Elastic Compute Cloud (Amazon EC2) GPU cases, offering full management over employee mannequin choice and batch throughput optimization. For productionizing a supervisor agent with its specialised instruments and dynamic data base, Bedrock AgentCore supplies the runtime scalability, reminiscence administration, and observability wanted to deploy self-learning programs reliably at scale.
Our supervisor agent integrates with Amazon’s intensive Choice and Catalog Methods. The above diagram is a simplified view displaying the important thing options of the agent and a few of the AWS companies that make it attainable. Product information flows via generator-evaluator staff (Amazon EC2 and Amazon Bedrock Runtime), with agreements saved straight and disagreements routed to a supervisor agent (Bedrock AgentCore). The training aggregator and reminiscence supervisor make the most of Amazon DynamoDB for the data base, with learnings injected again into employee prompts. Human overview (Amazon Easy Queue Service (Amazon SQS)) and observability (Amazon CloudWatch) full the structure. Manufacturing implementations will probably require extra elements for scale, reliability, and integration with current programs.
However how did we arrive at this structure? The important thing perception got here from an surprising place.
The perception: Turning disagreements into alternatives
Our perspective shifted throughout a debugging session. When a number of smaller fashions (equivalent to Nova Lite) disagreed on product attributes—decoding the identical specification otherwise based mostly on how they understood technical terminology—we initially noticed this as a failure. However the information instructed a distinct story: merchandise the place our smaller fashions disagreed correlated with circumstances requiring extra guide overview and clarification. When fashions disagreed, these had been exactly the merchandise that wanted extra investigation. The disagreements had been surfacing studying alternatives, however we couldn’t have engineers and scientists deep-dive on each case. The supervisor agent does this mechanically at scale. And crucially, the objective isn’t simply to find out which mannequin was proper—it’s to extract learnings that assist stop related disagreements sooner or later. That is the important thing to environment friendly scaling. Disagreements don’t simply come from AI staff at inference time. Put up-inference, sellers categorical disagreement via itemizing updates and appeals—alerts that our unique extraction may need missed necessary context. Prospects disagree via returns and destructive evaluations, typically indicating that product data didn’t match expectations. These post-inference human alerts feed into the identical studying pipeline, with the supervisor investigating patterns and producing learnings that assist stop related points throughout future merchandise. We discovered a candy spot: attributes with average AI employee disagreement charges yielded the richest learnings—excessive sufficient to floor significant patterns, low sufficient to point solvable ambiguity. When disagreement charges are too low, they sometimes mirror noise or basic mannequin limitations somewhat than learnable patterns—for these, we think about using extra succesful staff. When disagreement charges are too excessive, it alerts that employee fashions or prompts aren’t but mature sufficient, triggering extreme supervisor calls that undermine the effectivity positive aspects of the structure. These thresholds will range by job and area; the secret is figuring out your personal candy spot the place disagreements characterize real complexity value investigating, somewhat than basic gaps in employee functionality or random noise.
Deep dive: The way it works
On the coronary heart of our system are a number of light-weight employee fashions working in parallel—some as turbines extracting attributes, others as evaluators assessing these extractions. These staff may be carried out in a non-agentic approach with mounted inputs, making them batch-friendly and scalable. The generator-evaluator sample creates productive rigidity, conceptually just like the productive rigidity in generative adversarial networks (GANs), although our strategy operates at inference time via prompting somewhat than coaching. We explicitly immediate evaluators to be important, instructing them to scrutinize extractions for ambiguities, lacking context, or potential misinterpretations. This adversarial dynamic surfaces disagreements that characterize real complexity somewhat than letting ambiguous circumstances move via undetected. When the generator and evaluator agree, we have now excessive confidence within the outcome and course of it at minimal computational price. This consensus path handles most product attributes. Once they disagree, we’ve recognized a case value investigating—triggering the supervisor to resolve the dispute and extract reusable learnings.
Our structure treats disagreement as a common studying sign. At inference time, worker-to-worker disagreements catch ambiguity. Put up-inference, vendor suggestions catches misalignments with intent and buyer suggestions catches misalignments with expectations. The three channels feed the supervisor, which extracts learnings that enhance accuracy throughout the board. When staff disagree, we invoke a supervisor agent—a extra succesful mannequin that resolves the dispute and investigates why it occurred. The supervisor determines what context or reasoning the employees lacked, and these insights turn into reusable learnings for future circumstances. For instance, when staff disagreed about utilization classification for a product based mostly on sure technical phrases, the supervisor investigated and clarified that these phrases alone had been inadequate—visible context and different indicators wanted to be thought of collectively. The supervisor generated a studying about learn how to correctly weight totally different alerts for that product class. This studying instantly up to date our data base, and when injected into employee prompts for related merchandise, helped stop future disagreements throughout hundreds of things. Whereas the employees may theoretically be the identical mannequin because the supervisor, utilizing smaller fashions is essential for effectivity at scale. The architectural benefit emerges from this asymmetry: light-weight staff deal with routine circumstances via consensus, whereas the extra succesful supervisor is invoked solely when disagreements floor high-value studying alternatives. Because the system accumulates learnings and disagreement charges drop, supervisor calls naturally decline—effectivity positive aspects are baked straight into the structure. This worker-supervisor heterogeneity additionally allows richer investigation. As a result of supervisors are invoked selectively, they will afford to drag in extra alerts—buyer evaluations, return causes, vendor historical past—that may be impractical to retrieve for each product however present essential context when resolving complicated disagreements. When these alerts yield generalizable insights about how clients need product data introduced—which attributes to focus on, what terminology resonates, learn how to body specs—the ensuing learnings profit future inferences throughout related merchandise with out retrieving these resource-intensive alerts once more. Over time, this creates a suggestions loop: higher product data results in fewer returns and destructive evaluations, which in flip displays improved buyer satisfaction.
The data base: Making learnings scalable
The supervisor investigates disagreements on the particular person product degree. With tens of millions of things to course of, we want a scalable technique to rework these product-specific insights into reusable learnings. Our aggregation technique adapts to context: high-volume patterns get synthesized into broader learnings, whereas distinctive or important circumstances are preserved individually. We use a hierarchical construction the place a big language mannequin (LLM)-based reminiscence supervisor navigates the data tree to position every studying. Ranging from the foundation, it traverses classes and subcategories, deciding at every degree whether or not to proceed down an current path, create a brand new department, merge with current data, or exchange outdated data. This dynamic group permits the data base to evolve with rising patterns whereas sustaining logical construction. Throughout inference, staff obtain related learnings of their prompts based mostly on product class, mechanically incorporating area data from previous disagreements. The data base additionally introduces traceability—when an extraction appears incorrect, we will pinpoint precisely which studying influenced it. This shifts auditing from an unscalable job to a sensible one: as an alternative of reviewing a pattern of tens of millions of outputs—the place human effort grows proportionally with scale—groups can audit the data base itself, which stays comparatively mounted in dimension no matter inference quantity. Area specialists can straight contribute by including or refining entries, no retraining required. A single well-crafted studying can instantly enhance accuracy throughout hundreds of merchandise. The data base bridges human experience and AI functionality, the place automated learnings and human insights work collectively.
Classes realized and greatest practices
When this self-learning structure works greatest:
- Excessive-volume inference the place enter variety drives compounded studying
- High quality-critical functions the place consensus supplies pure high quality assurance
- Evolving domains with new patterns and terminology consistently rising
It’s much less appropriate for low-volume situations (inadequate disagreements for studying) or use circumstances with mounted, unchanging guidelines.
Crucial success elements:
- Defining disagreements: With a generator-evaluator pair, disagreement happens when the evaluator flags the extraction as needing enchancment. With a number of staff, scale thresholds accordingly. The secret’s sustaining productive rigidity between staff. If disagreement charges fall outdoors the productive vary (too low or too excessive), take into account extra succesful staff or refined prompts.
- Monitoring studying effectiveness: Disagreement charges should lower over time—that is your major well being metric. If charges keep flat, verify data retrieval, immediate injection, or evaluator criticality.
- Information group: Construction learnings hierarchically and maintain them actionable. Summary steerage doesn’t assist; particular, concrete learnings straight enhance future inferences.
Frequent pitfalls
- Specializing in price over intelligence: Value discount is a byproduct, not the objective
- Rubber-stamp evaluators: Evaluators that merely approve generator outputs gained’t floor significant disagreements—immediate them to actively problem and critique extractions
- Poor studying extraction: Supervisors should establish generalizable patterns, not simply repair particular person circumstances
- Information rot: With out group, learnings turn into unsearchable and unusable
The important thing perception: deal with declining disagreement charges as your north star metric—they present the system is actually studying.
Deployment methods: Two approaches
- Study-then-deploy: Begin with primary prompts and let the system be taught aggressively in a pre-production setting. Area specialists then audit the data base—not particular person outputs—to ensure realized patterns align with desired outcomes. When authorised, deploy with validated learnings. That is ideally suited for brand new use circumstances the place you don’t but know what good appears like—disagreements assist uncover the proper patterns, and data base auditing permits you to form them earlier than manufacturing.
- Deploy-and-learn: Begin with refined prompts and good preliminary high quality, then repeatedly enhance via ongoing studying in manufacturing. This works greatest for well-understood use circumstances the place you possibly can outline high quality upfront however nonetheless wish to seize domain-specific nuances over time.
Each approaches use the identical structure—the selection is determined by whether or not you’re exploring new territory or optimizing acquainted floor.
Conclusion
What began as an experiment in catalog enrichment revealed a basic reality: AI programs don’t must be frozen in time. By embracing disagreements as studying alerts somewhat than failures, we’ve constructed an structure that accumulates area data via precise utilization. We watched the system evolve from generic understanding to domain-specific experience. It realized industry-specific terminology. It found contextual guidelines that change throughout classes. It tailored to necessities no pre-trained mannequin would encounter—all with out retraining, via learnings saved in a data base and injected again into employee prompts. For groups operationalizing related architectures, Amazon Bedrock AgentCore provides purpose-built capabilities:
- AgentCore Runtime handles fast consensus selections for routine circumstances whereas supporting prolonged reasoning when supervisors examine complicated disagreements
- AgentCore Observability supplies visibility into which learnings drive affect, serving to groups refine data propagation and keep reliability at scale
The implications prolong past catalog administration. Excessive-volume AI functions may benefit from this course of—and the power of Amazon Bedrock to entry various fashions makes this structure simple to implement. The important thing perception is that this: we’ve shifted from asking “which mannequin ought to we use?” to “how can we construct programs that be taught our particular patterns? “Whether or not you learn-then-deploy for brand new use circumstances or deploy-and-learn for established ones, the implementation is simple: begin with staff suited to your job, select a supervisor, and let disagreements drive studying. With the proper structure, each inference can turn into a chance to seize area data. That’s not simply scaling—that’s constructing institutional data into your AI programs.
Acknowledgement
This work wouldn’t have been attainable with out the contributions and assist from Ankur Datta (Senior Principal Utilized Scientist – chief of science in On a regular basis Necessities Shops), Zhu Cheng (Utilized Scientist), Xuan Tang (Software program Engineer), Mohammad Ghasemi (Utilized Scientist). We sincerely recognize the contributions in designs, implementations, quite a few fruitful brain-storming periods, and all of the insightful concepts and recommendations.
Concerning the authors
Tarik Arici is a Principal Scientist at Amazon Choice and Catalog Methods (ASCS), the place he pioneers self-learning generative AI programs design for catalog high quality enhancement at scale. His work focuses on constructing AI programs that mechanically accumulate area data via manufacturing utilization—studying from buyer evaluations and returns, vendor suggestions, and mannequin disagreements to enhance high quality whereas lowering prices. Tarik holds a PhD in Electrical and Pc Engineering from Georgia Institute of Expertise.
Sameer Thombare is a Senior Product Supervisor at Amazon with over a decade of expertise in Product Administration, Class/P&L Administration throughout various industries, together with heavy engineering, telecommunications, finance, and eCommerce. Sameer is keen about creating repeatedly enhancing closed-loop programs and leads strategic initiatives inside Amazon Choice and Catalog Methods (ASCS) to construct a complicated self-learning closed-loop system that synthesize alerts from clients, sellers, and provide chain operations to optimize outcomes. Sameer holds an MBA from the Indian Institute of Administration Bangalore and an engineering diploma from Mumbai College.
Amin Banitalebi acquired his PhD within the Digital Media on the College of British Columbia (UBC), Canada, in 2014. Since then, he has taken varied utilized science roles spanning over areas in laptop imaginative and prescient, pure language processing, advice programs, classical machine studying, and generative AI. Amin has co-authored over 90 publications and patents. He’s at the moment an Utilized Science Supervisor in Amazon On a regular basis Necessities.
Puneet Sahni is a Senior Principal Engineer at Amazon Choice and Catalog Methods (ASCS), the place he has spent over 8 years enhancing the completeness, consistency, and correctness of catalog information. He focuses on catalog information modeling and its utility to enhancing Promoting Associate and buyer experiences, whereas utilizing ML/DL and LLM-based enrichment to drive enhancements in catalog information high quality.
Erdinc Basci joined Amazon in 2015 and brings over 23 years of expertise {industry} expertise. At Amazon, he has led the evolution of Catalog system architectures—together with ingestion pipelines, prioritized processing, and visitors shaping—in addition to catalog information structure enhancements equivalent to segmented provides, product specs for manufacture-on-demand merchandise, and catalog information experimentation. Erdinc has championed a hands-on efficiency engineering tradition throughout Amazon companies unlocking $1B+ annualized price financial savings and 20%+ latency wins throughout core Shops companies. He’s at the moment centered on enhancing generative AI utility efficiency and GPU effectivity throughout Amazon. Erdinc holds a BS in Pc Science from Bilkent College, Turkey, and an MBA from Seattle College, US.
Mey Meenakshisundaram is a Director in Amazon Choice and Catalog Methods, the place he leads progressive GenAI options to ascertain Amazon’s worldwide catalog because the best-in-class supply for product data. His workforce pioneers superior machine studying strategies, together with multi-agent programs and huge language fashions, to mechanically enrich product attributes and enhance catalog high quality at scale. Excessive-quality product data within the catalog is important for delighting clients find the proper merchandise, empowering promoting companions to checklist their merchandise successfully, and enabling Amazon operations to scale back guide effort.
