Selected Case Studies

From Technical Possibility to Measurable Business Capability

The technologies, industries, and operating environments have varied. The underlying pattern has remained consistent: identify where intelligence can change the business, build the capability, and connect it to measurable value.

These are outcomes from in-house executive roles at prior employers, not Rhiza Advisory client engagements. Companies are named where the work is a matter of record and described by type otherwise; figures are stated as approximations with the contribution accurately bounded.

AI-Enabled Product Growth

Connected-products company · in-house executive role · consumer and commercial water-management devices

Turn connected-device data into an intelligent customer capability and a scalable commercial model.

The situation

The devices were already deployed and already generating telemetry. The data existed; the commercial capability did not. Device intelligence was treated as a product feature rather than as recurring-revenue infrastructure, so the telemetry accumulated without becoming something a customer would pay for on an ongoing basis.

Constraints

  • Production intelligence had to work reliably across a live installed base, not just a controlled pilot.
  • Model output had to be reliable enough to act on, because a false leak alert carries a real service cost and a real trust cost.
  • The capability had to ship into an existing product and an existing customer base rather than a greenfield build.

Role held

Executive owner of the data and AI capability, working across product, engineering, and commercial leadership.

Decisions that mattered

  • Prioritized production reliability over model sophistication, since the commercial model depended on alerts customers would trust and act on.
  • Built the capability as recurring-revenue infrastructure rather than as a feature, so the value compounded with the installed base.

What changed

Built production machine-learning capabilities supporting leak detection, product intelligence, and recurring customer value.

Measured outcome

Helped support approximately 12× estimated growth in annual revenue run rate during this period.

Why this transfers

For connected products, the win is not the model. It is turning device data into a capability customers keep paying for. Built as recurring-revenue infrastructure rather than a feature, product intelligence compounds with the installed base into enterprise value an acquirer can underwrite.

Predictive Risk & Algorithmic Trading Systems

Morgan Stanley, formerly Dean Witter Reynolds · U.S. Treasury primary dealer · fixed-income trading and risk

Turn a fixed-income trading floor into a predictive, algorithmic decision system.

The situation

At Morgan Stanley, formerly Dean Witter Reynolds, the fixed-income operation managed positions across U.S. Treasuries, agencies, corporate and municipal bonds, mortgage-backed securities, money markets, zero-coupon securities, and repo financing.

Risk management was fragmented and largely manual. Traders lacked a common system for valuing positions and understanding exposure across the floor. Many less-liquid securities did not trade frequently enough to provide reliable market prices, making it difficult to quantify risk consistently or determine the right response to changes in interest rates, the yield curve, duration, and convexity.

Constraints

  • The system had to manage risk across the entire fixed-income operation, not a single desk or asset class.
  • Securities without reliable live market prices still had to be valued accurately enough to support real trading and hedging decisions.
  • The output had to prescribe an action, not merely aggregate positions or report risk after the fact.
  • Traders needed to retain judgment and control while the system automated the analytical work behind pricing, exposure, hedge selection, and trading opportunities.
  • Everything had to operate within the market-data, computing, and execution constraints of the 1990s.

Role held

Designed and built the predictive risk-management platform for the Dean Witter fixed-income operation within Morgan Stanley, working directly with traders and trading-floor leadership.

The platform priced illiquid securities, consolidated exposure across asset classes, quantified portfolio risk, and recommended the securities and quantities required to hedge yield-curve, duration, and convexity exposure.

Also designed and built an algorithmic U.S. Treasury bond arbitrage system that identified and drove stripping and reconstitution opportunities using real-time predictive pricing.

Decisions that mattered

  • Replaced fragmented, trader-by-trader risk management with a shared operating system for the entire fixed-income floor.
  • Used predictive models to price securities that could not be reliably valued from active market quotes.
  • Made the platform prescriptive by recommending both the quantity and specific securities required to offset risk.
  • Treated hedge selection as an optimization problem across yield-curve, duration, and convexity exposure, not simply a reporting exercise.
  • Applied the same predictive intelligence to U.S. Treasury bond arbitrage, continuously comparing the value of the underlying bond with its component STRIPS and identifying opportunities in both directions.
  • Kept traders at the point of execution while allowing algorithms to determine the pricing signals, trade structure, and recommended action.

What changed

Moved the fixed-income operation from largely manual and fragmented risk management to a common predictive and prescriptive platform covering the full trading floor.

Traders gained a shared view of positions, modeled prices, exposure, and recommended hedges across cash securities and repo financing. Risk could be evaluated consistently across asset classes, including securities without active market prices.

The related algorithmic trading system converted real-time Treasury-market data into actionable arbitrage decisions: predicting the value of illiquid STRIPS, comparing those values with the reconstituted bond, and identifying when to strip or reconstitute securities.

These systems did not merely display information. They predicted values, quantified risk, prescribed actions, and changed how trading decisions were made.

Measured outcome

Brought an approximately $10 billion gross fixed-income book under automated, predictive risk management across the entire Dean Witter fixed-income operation within Morgan Stanley.

The algorithmic U.S. Treasury bond arbitrage system supported approximately $300 million in cumulative face value stripped and reconstituted through model-generated pricing and trading decisions.

The approximately $10 billion figure is anchored to the company’s audited 1996 merger pro-forma disclosures. It represents gross long and short fixed-income inventory together with repo and reverse-repo financing, not assets under management or a net directional position.

The approximately $300 million figure represents estimated cumulative face value processed through the U.S. Treasury bond arbitrage strategy, based on firsthand operating knowledge of the system and its trading activity.

Why this transfers

When data is incomplete, conditions are changing, and decisions carry material financial consequences, a dashboard is not enough.

The advantage comes from a system that can predict what cannot be directly observed, quantify exposure, recommend the next action, and preserve human judgment at the point of decision. That is the same operating pattern behind modern AI-enabled products, risk systems, supply chains, and other complex businesses where intelligence must translate into action before the opportunity, or the risk, moves.

AI Economics at Scale

Connected-products company · in-house executive role · consumer and commercial water-management devices

Make production AI economically viable before growth turns usage into a scaling tax.

The situation

The company’s connected devices were generating the data required to deliver intelligent leak detection and product insights. But one of the AI pipelines relied on costly centralized batch processing, creating a recurring expense for every active device.

At an early installed base, the cost appeared manageable. At consumer-product scale, however, the architecture threatened the viability of the emerging business model by causing AI operating expense to rise nearly in line with customer adoption.

Constraints

  • The cost multiplied with every device added, turning customer growth into recurring infrastructure expense.
  • The architecture had to support an installed product and customer base rather than a controlled pilot or greenfield system.
  • Cost reduction could not come at the expense of reliable product intelligence or the customer experience.
  • The solution needed to support continued growth without requiring cloud expense to rise at the same rate as the installed base.

Role held

Executive owner of the data and AI capability, reporting to the CEO and working across data science, engineering, product, and commercial leadership.

Led the effort to treat inference economics as a product and business-model issue, not merely a technical infrastructure concern.

Decisions that mattered

  • Made cost per active device a first-class product metric because it determined whether recurring revenue would remain attractive as adoption grew.
  • Rejected an architecture in which AI operating expense scaled nearly linearly with the installed base.
  • Shifted the pipeline from costly centralized batch processing toward more efficient live/on-device inference.
  • Optimized for the economics of the complete product, not simply model accuracy or technical sophistication in isolation.

What changed

Re-architected the AI pipeline so that product intelligence could be delivered at a fraction of the previous recurring cost.

The new approach reduced dependence on heavy centralized processing, materially improved per-device economics, and created an architecture better aligned with the company’s connected-product and subscription growth model.

AI moved from being a feature whose cost threatened to compound with adoption to a capability whose economics improved the viability of growth.

Measured outcome

Reduced monthly AI operating cost per device by approximately 91%, bringing the redesigned architecture to less than 9% of the prior cost structure.

Because the savings scaled with every active device, the change materially improved unit economics and prevented AI operating expense from increasing nearly in line with customer growth. At equivalent deployment volumes, the redesigned architecture reduced annualized operating costs by approximately 91% while preserving the production intelligence required by the product.

Figures are based on the approximate monthly processing cost per active device before and after the architecture change. Scale projections apply those unit costs consistently across the stated installed-base scenarios and are intended to demonstrate the economic effect of architecture at growth-stage volumes.

Why this transfers

For AI-enabled products, inference cost is part of the business model, not an implementation detail.

A system can perform well technically and still destroy value if every new customer adds disproportionate recurring expense. Production AI has not truly reached scale until its reliability, architecture, and unit economics can survive adoption.

Discuss Your AI Priorities