# Raja Jeevan Kumar Maduri

> Generated by All in One SEO v5.0.2.1, this is an llms.txt file, used by LLMs to index the site.

Clear product decisions for complicated commerce systems

## Sitemaps

- [XML Sitemap](https://rajajeevan.com/sitemap.xml): Contains all public & indexable URLs for this website.

## Pages

- [Home](https://rajajeevan.com/)
- [Sitemap](https://rajajeevan.com/sitemap/): Every public page on rajajeevan.com: the core pages, all portfolio entries, the writing and the reference resources.
- [FAQs about Raja](https://rajajeevan.com/faq/): Over 100 short answers about Raja Jeevan Kumar Maduri: his background and roles, retail systems and AI expertise, advisory services, published work and how to contact him.
- [Brand](https://rajajeevan.com/brand/)
- [Metric catalogue for vision checkout and unattended retail: reference design](https://rajajeevan.com/metric-catalogue-for-vision-checkout-and-unattended-retail-reference-design/): Deliverable | Retail technology | KPI framework | Reference design, no targets or resultsHow to use this catalogueThe author contributed domain expertise to solutions of this kind at a medium-scale, product-based company based out of Dallas and Coimbatore. This catalogue is a reference design. It lists measures that a retailer or vendor can choose from
- [Writing](https://rajajeevan.com/writing/)

## Portfolio Entries

- [Walk-out retail is a shrinkage and labour bet, not a convenience feature: a reference design for autonomous stores and smart coolers](https://rajajeevan.com/portfolio/walk-out-retail-is-a-shrinkage-and-labour-bet-not-a-convenience-feature-a-reference-design-for-autonomous-stores-and-smart-coolers/): Case study | Retail technology | Unattended retail | Problem, opportunity and reference design Key findings Unattended retail only pays when it removes cost the store actually carries, such as labour hours and shrink. Convenience alone does not fund it. The format is a spectrum. A smart cooler, a micro-market and a full walk-out store
- [The warehouse needs a guidance layer](https://rajajeevan.com/portfolio/the-warehouse-needs-a-guidance-layer/): A warehouse management system can record a pick after it happens. What would it take to catch a mismatch while a person can still correct it? This concept sets out an observe, compare, guide and confirm pattern, with human control and field testing still open.
- [Your POS migration is also a data-ownership decision](https://rajajeevan.com/portfolio/your-pos-migration-is-also-a-data-ownership-decision/): The sales receipt is the visible part of a checkout. Behind it sit products, prices, tax rules, tender types, staff permissions and a history of changes.
- [Product ontology before product AI](https://rajajeevan.com/portfolio/product-ontology-before-product-ai/): When product categories and attributes disagree, adding GenAI can make catalog errors harder to spot. This entry shows my approach to a product ontology: organize the relationships first, then design enrichment and integration. It does not claim a verified uplift in discovery or revenue.
- [A price change is not finished at headquarters](https://rajajeevan.com/portfolio/a-price-change-is-not-finished-at-headquarters/): Picture a shopper holding an item at the shelf. The label says one price. The till says another. The shopper does not care which system was updated first. The store has broken a promise.
- [Self-checkout fails on recognition and exceptions, not on the screen: a reference design for vision-based checkout](https://rajajeevan.com/portfolio/self-checkout-fails-on-recognition-and-exceptions-not-on-the-screen-a-reference-design-for-vision-based-checkout/): Case study | Retail technology | Computer vision | Problem, opportunity and reference design Key findings Retailers have already scaled self-checkout. The open problem is losses and exceptions at the lane, not whether shoppers will use it. A vision-based tray checkout replaces scanning with recognition. That moves risk from the shopper’s hands to the model’s
- [Structure before intelligence: how a GenAI product ontology turned catalogue maintenance into an exception process](https://rajajeevan.com/portfolio/structure-before-intelligence-how-a-genai-product-ontology-turned-catalogue-maintenance-into-an-exception-process/): Case study | Retail and commerce systems | Generative AI | Engagement period: March 2024 to March 2025 Key findings Product data fails at the structure layer first. Search, discovery and categorisation errors trace back to an inconsistent taxonomy and attribute model, and every downstream system inherits them. Generative AI helps most when it works
- [A smart mirror for the morning: a gesture-driven computer vision concept and build blueprint](https://rajajeevan.com/portfolio/smart-mirror-gesture-computer-vision-concept-and-build-blueprint/): A proposed concept and build blueprint for a gesture-driven smart mirror that helps a user get ready for the day. It covers the experience, computer vision, hardware, software, network, systems and operations, with every claim tagged by evidence level. It is a concept, not a deployment.
- [The Harness Council: an operating philosophy for AI knowledge work](https://rajajeevan.com/portfolio/harness-council-operating-philosophy/): A first-person experiment in AI knowledge work: one orchestrator, a council of specialist harnesses, and evidence before every done. Every claim is labelled Reported, Observed, Design or Roadmap. Kept current as the experiment changes.
- [Before the first drill hole: a retrieval-augmented knowledge system for mineral exploration](https://rajajeevan.com/portfolio/before-the-first-drill-hole-a-retrieval-augmented-knowledge-system-for-mineral-exploration/): Case study | Mining and resources | Retrieval-augmented generation | Reference design Key findings Exploration decisions rest on data that already exists. Geological maps, geophysical and geochemical surveys, earlier drilling and tenure records are mostly collected before a rig arrives, and they are scattered across formats, decades and organisations. Retrieval-augmented generation fits this problem because
- [Hermes Harness: designing for lower operational overhead](https://rajajeevan.com/portfolio/hermes-harness-designing-for-lower-operational-overhead/): A detailed experiment in coordinating specialist AI profiles, choosing structured execution paths and checking outcomes. Implementation changes are attributed to Hermes's own execution report, not independently verified; performance estimates are excluded.
- [From fifteen use cases to a measurable catalogue: how a retailer's camera estate becomes a metrics layer](https://rajajeevan.com/portfolio/from-fifteen-use-cases-to-a-measurable-catalogue-how-a-retailers-camera-estate-becomes-a-metrics-layer/): Case study | Retail operations | Video analytics | Solution design for an existing store estateKey findingsA video analytics proposal is only as persuasive as its metric catalogue. Buyers approve numbers they can read in a weekly review, not capabilities they cannot price.The value sits in the join. A camera count means little alone. A
- [Autonomy with gates: designing an AI software factory for a regulated legal-technology platform](https://rajajeevan.com/portfolio/autonomy-with-gates-designing-an-ai-software-factory-for-a-regulated-legal-technology-platform/): Case study | Engineering operating model | Agentic AI and governance | Design programme, 2026 Key findings Agentic coding raises throughput, but throughput without governance moves risk faster than value. The design question is where humans stay in the loop, not whether to use agents. A three-plane model separates concerns cleanly: a control plane that
- [The cameras are already paid for: pricing the hidden value of a retailer's surveillance estate](https://rajajeevan.com/portfolio/the-cameras-are-already-paid-for-pricing-the-hidden-value-of-a-retailers-surveillance-estate/): Case study | Retail operations | Computer vision and analytics | Solution proposal, September to October 2021 Key findings A large installed camera base is a sunk cost that can be turned into an operating-data asset without replacing a single device. The winning move against an incumbent was specificity. A catalogue of about 150 metrics,
- [The hidden wall in every computer-vision business case is data: a reference design for synthetic data and assisted annotation](https://rajajeevan.com/portfolio/the-hidden-wall-in-every-computer-vision-business-case-is-data-a-reference-design-for-synthetic-data-and-assisted-annotation/): Case study | Practical AI | Computer vision data operations | Problem, opportunity and reference design Key findings In retail computer vision, the model is rarely the bottleneck. Getting enough correctly labelled images of the right situations is. The catalogue changes constantly, so a labelled set is never finished. Data operations need a standing process,
- [One platform, many markets: a proposed design for global sales and operations planning at a quick-service chain](https://rajajeevan.com/portfolio/one-platform-many-markets-a-proposed-design-for-global-sales-and-operations-planning-at-a-quick-service-chain/): Case study | Quick-service restaurants | Supply chain planning | Proposed solution for a multi-market deployment (response prepared August 2024)
- [Designing a five-tier inference fallback for an AI agent](https://rajajeevan.com/portfolio/designing-a-five-tier-inference-fallback-for-an-ai-agent/): An agent needs a deliberate path when its first inference route slows down or fails. This five-tier architecture experiment maps the handoff choices and what should be observed at each step. It is a design record, not proof of production uptime, low cost or answer quality.
- [A personal knowledge-curation loop](https://rajajeevan.com/portfolio/a-personal-knowledge-curation-loop/): A personal AI workflow can screen material against interests the reader declares, then prepare a short brief for review. This entry explains the curation loop I am building and the questions it still needs to answer. It makes no claim of measured time savings or automatic improvement.
- [Frontline AI has to help the person doing the work](https://rajajeevan.com/portfolio/frontline-ai-has-to-help-the-person-doing-the-work/): Picture a colleague carrying out a task on a busy floor. A system notices a possible mistake. What should happen next: an alert to a manager, a useful prompt to the colleague, or nothing until a person checks the evidence?
- [Sell the next useful action, not another automation canvas](https://rajajeevan.com/portfolio/sell-the-next-useful-action-not-another-automation-canvas/): A small-business owner does not wake up wanting a workflow engine. They want fewer missed appointments, clearer order updates or a way to follow up without losing the thread.
- [A learning app should make practice easier, not guilt louder](https://rajajeevan.com/portfolio/a-learning-app-should-make-practice-easier-not-guilt-louder/): A student opens a learning app after missing several days. The streak is gone. The backlog is longer. The app has measured absence, but has it helped the student begin again?
- [The knowledge bot that needs an owner before it needs a model](https://rajajeevan.com/portfolio/the-knowledge-bot-that-needs-an-owner-before-it-needs-a-model/): Imagine asking a company knowledge bot how a product works. It answers confidently. Then an engineer points out that the document it used belongs to an older release.
- [A group-buying app has to earn the group's trust](https://rajajeevan.com/portfolio/a-group-buying-app-has-to-earn-the-groups-trust/): Picture a housing society agreeing to buy an essential together. The price looks attractive. Then the questions begin: who chose the vendor, who agreed to the quantity, and what happens if a household changes its mind?
- [A hospital surge forecast is only useful if someone can act on it](https://rajajeevan.com/portfolio/a-hospital-surge-forecast-is-only-useful-if-someone-can-act-on-it/): Imagine a hospital receiving an early warning that demand may rise. The warning lands in a dashboard. Staffing, supplies and patient-flow decisions stay unchanged.
- [A video-intelligence layer for store operations](https://rajajeevan.com/portfolio/a-video-intelligence-layer-for-store-operations/): Store cameras can capture movement and queues, but footage alone does not tell a team what to do. I designed a video-intelligence concept connecting potential signals to operational decisions, integration work and a measurement plan. This is solution design, not a production deployment.
- [The proposal that began with the cameras already in place](https://rajajeevan.com/portfolio/the-proposal-that-began-with-the-cameras-already-in-place/): A retailer already had cameras, but no clear way to turn footage into operating decisions. This proposal framed video analytics around questions a store team could act on, using existing systems. It was a proposal, not a deployment or measured result.

## Writing

- [Between scan and settlement: the checkout plumbing nobody sees](https://rajajeevan.com/writing/between-scan-and-settlement-the-checkout-plumbing-nobody-sees/): What happens between the scan and the settlement at a checkout, where it breaks, and how to design for the unknown: references, explicit states, safe retries, refunds and reconciliation.
- [When the till disagrees: POS exception handling a store team can run](https://rajajeevan.com/writing/when-the-till-disagrees-pos-exception-handling-a-store-team-can-run/): A store till disagrees with the shopper many times a day. How to design exception handling that a store team can run without calling head office: five questions every exception must answer, override rights, offline mode and routing to owners.
- [The most useful decision an AI agent can make is to stop](https://rajajeevan.com/writing/the-most-useful-decision-an-ai-agent-can-make-is-to-stop/): My multi-agent setup uses evidence gates and human review. That is a practical lesson about calibrated decisions, not proof that the agents run reinforcement learning.
- [Building a product ontology: a practical how-to for retail teams](https://rajajeevan.com/writing/building-a-product-ontology-a-practical-how-to-for-retail-teams/): A practical how-to for building a retail product ontology: classes, attributes, relationships, a model that proposes while people approve, and turning catalogue upkeep into an exception process. Companion to the ontology case study.
- [Why the stock count never matches: a field guide to reconciliation loops](https://rajajeevan.com/writing/why-the-stock-count-never-matches-a-field-guide-to-reconciliation-loops/): Why a store's stock count and the shelf never agree, and a loop that keeps the gap small: where drift comes from, how to detect, diagnose, decide and correct, and how to run cycle counts that find something.