Our Work

A major retailer with 2,000+ stores had inconsistent demand forecasts across merchandising, supply chain, and finance. Each function used its own models, leading to overstocks and stockouts.

Unified demand forecasting platform covering 200K SKUs across stores and online channels, producing internally consistent forecasts at the SKU-store-day level. Scaled team from 5 to 45.

Unified forecasting across historically siloed functions. The shipped model significantly outperformed the industry-standard alternative used by most retailers. $3-4B in free cash flow unlocked.

Retailers need to understand which products customers treat as interchangeable. Traditional category hierarchies miss context-dependent relationships.

Learned product and store representations that capture substitutability, complementarity, and customer preference from transaction data, feeding into forecasting, pricing, and assortment.

Context-aware product relationships that go beyond static category hierarchies.

Companies growing through acquisitions inherit poorly located warehouses. Inventory arrives too early (tying up capital) or too late (causing stockouts).

Simulation-based network design optimizing warehouse locations for lead times, transport costs, labor, and disruption risk. Combined with real-time inventory control across the full multi-tier network.

End-to-end optimization from strategic network design down to real-time inventory movement decisions.

Inventory counts are inaccurate. Cycle counts are expensive and infrequent. Phantom stockouts cause lost sales.

Continuous inventory estimation by triangulating multiple data sources (sales transactions, inbound/outbound scans, camera feeds, RFID) without requiring manual counts.

Probabilistic state estimation replacing expensive manual cycle counts with continuous automated accuracy.

Laboratories order excess consumables to avoid running out, but face strict overnight storage limits. Over-ordering wastes transport labor and creates compliance risk.

Forecasting model predicting daily consumption per lab from historical delivery and return patterns, enabling just-in-time delivery of the correct volume.

Demand forecasting adapted for regulated lab environments with hard storage constraints.

Supply chain disruptions exposed the fragility of single-source procurement. Companies need early warning of supplier instability.

Continuous monitoring of supplier health with automated risk playbooks that trigger contingency actions based on real-time risk assessments.

Proactive risk management replacing reactive firefighting with automated early-warning and response.

A major retailer had failed three times to deploy clearance pricing optimization. Merchants had lost trust in AI-driven pricing.

Sequential decision problem solved with multiple optimization approaches. Embedded directly into merchant workflows so users see recommendations alongside projected financial impact.

Succeeded where three prior attempts failed by embedding the system in existing workflows. $100M+ profit impact.

True incremental lift from promotions was hard to isolate. Merchants ran promotions on gut feel. Supply chain had no advance visibility into promotional periods.

Promotional lift estimation isolating the true incremental effect, integrated with demand forecasting so promotional plans automatically adjust supply chain planning.

Connected promotional planning and supply chain preparation into a single system. When a merchant plans a promotion, inventory positioning adjusts automatically.

A major grocer needed to scale its advertising business with experimentation, forecasting, and optimization engines.

Analytics backbone: experimentation engines to measure ad effectiveness, forecasting for ad placement inventory, and optimization to match advertiser budgets with customer segments.

Accelerated Retail Media to 30% CAGR at 50% margin.

Generic campaigns sent the same offer to every customer, wasting budget.

Personalization engines determining the right product, offer, and message for each customer at each moment, factoring in purchase history, category affinity, price sensitivity, and inventory position.

Real-time personalization that combines customer intelligence with operational constraints like inventory.

During key retail moments, generic campaigns waste budget. Each customer has different preferences and price sensitivity.

Recommendation system matching customers to offers in real time, factoring in seasonality, promotional calendar, inventory position, and individual purchase history.

Seasonal-aware personalization that adapts to both the calendar and the individual customer.

Pharmacy and grocery operated as separate businesses. Customer data was siloed, meaning missed cross-sell opportunities.

Unified pharmacy and grocery customer data, connecting prescription behavior with grocery purchasing patterns. Lookalike modeling to identify cross-category opportunities.

Integrating two previously siloed data sources created a richer customer profile than either had independently.

Customers churn silently. By the time they are "lost," it is too late to intervene.

Unified customer view integrating purchase patterns, service interactions, digital behavior, and sentiment. Next-best-action engine recommending the optimal intervention per customer.

Proactive retention replacing reactive re-engagement with micro-targeted interventions.

ROAS attributes sales to a single touchpoint, missing the full customer journey across multiple channels.

Attribution model that decomposes the contribution of each marketing channel and touchpoint to a conversion, replacing last-click attribution with a full-journey model.

True multi-channel attribution enabling evidence-based marketing budget allocation.

Customers cannot describe visual attributes (drape, cut, pattern, texture) in keywords. Text search fails for fashion.

Computer vision with an attention mechanism that ignores background distractions and matches uploaded images against the product catalog.

Visual understanding replacing keyword search for inherently visual product categories.

Multiple teams were experimenting with GenAI independently without standards or governance, creating risk and duplicated effort.

25-person GenAI CoE owning the governed platform, guardrails, and standards. Production use cases across marketing, product content, search, call center, and store operations.

Moved GenAI from experimentation to governed production in under a year.

Marketing teams produce thousands of pieces of copy across dozens of audience segments. Manual copywriting does not scale.

LLM-powered copy generation integrated into the marketing workflow with brand voice and compliance guardrails enforced automatically.

Automated content generation with built-in brand consistency and compliance checks.

Marketing and e-commerce teams need product imagery variations for campaigns and A/B testing. Traditional photography is expensive and slow.

AI-powered image generation producing product imagery variations (backgrounds, styling, seasonal themes) from existing product photos.

On-demand product imagery at a fraction of the cost and time of traditional photography.

E-commerce product pages require titles, descriptions, and summaries for 200K+ SKUs. Manual creation is impossible at this scale.

Automated pipelines for SEO-optimized titles, descriptions from structured attributes, feature summaries, and review summaries distilling hundreds of reviews into highlights.

Automated content creation at a scale no human team could match, with SEO optimization built in.

E-commerce search returned irrelevant results for natural language queries because it relied on keyword matching.

LLM-enhanced search that understands query intent and maps natural language to products, combining keyword search with semantic understanding.

Natural language understanding replacing rigid keyword matching for product discovery.

Call center agents spend time searching for policies and crafting responses. Training is expensive and turnover is high.

Real-time AI assistance providing agents with suggested responses, policy lookups, and order context. New agents perform at experienced-agent levels from day one.

Eliminated the performance gap between new and experienced agents through real-time AI assistance.

Store and warehouse employees need quick answers to operational questions and currently call a manager or search manuals.

Chatbot on handheld devices answering operational questions from SOPs and policy documents. Bidirectional: employees also report local conditions back to central systems.

Closes the information loop between headquarters and the field, making it bidirectional.

Current e-commerce is search-and-filter. Customers with complex needs cannot express their requirements through keyword search.

AI agent that engages in dialogue, understands the customer's project, recommends products, builds bills of materials, schedules delivery, and processes payment.

Replaced search-and-filter with conversational product discovery for complex, multi-item purchases.

A contract research organization responds to hundreds of client questionnaires per year, each consuming 50-100+ person-hours. Questions are repetitive but worded differently.

Semantic matching against historical Q&A pairs to produce draft responses with confidence scores and source citations. Client-aware redaction prevents information leakage.

Semantic matching with client-aware redaction, ensuring one client's information never leaks into another's response.

Warehouse staff manually read vendor labels on reagent bottles. Different forms of the same compound have different identifiers, causing inventory confusion.

Vision AI for bilingual label capture, structured data extraction, storage recommendations, and a knowledge graph mapping compound names to all variant forms.

Knowledge graph resolving chemical compound identity across naming conventions and languages.

Over 1,000 chemists spend significant time on reports that must conform to patent-standard formats. Each uses a different style.

Locally deployed language model that reformats chemistry reports to the required standard template, preserving all scientific content while standardizing formatting.

On-premise LLM deployment for data-sovereign environments where cloud APIs are not permitted.

Drug discovery is a multi-year, multi-billion dollar process where most candidate molecules fail. Scientists need to predict whether a molecule will succeed before committing to expensive wet lab experiments.

ML models spanning the full drug design pipeline: PK-PD prediction, toxicity screening, synthesizability assessment, ADME property prediction, and molecular property optimization. Helps scientists prioritize which molecules to synthesize and test.

Applied AI across all stages of the drug design pipeline in a single advisory framework, letting scientists evaluate candidate molecules holistically.

Education technology is content-focused and generic. Students need specific learning paths based on their mastery gaps, not generic courses.

Platform mapping topic dependency graphs, tracking mastery per student per concept, classifying error patterns, and generating personalized learning paths. Teacher dashboard for class and per-student visibility.

Error classification (conceptual vs careless vs time pressure) enabling targeted remediation rather than generic review.

Made-to-order factories lose millions in scrap when production falls behind. Every piece is custom, so extras cannot be sold. Managers lack real-time visibility.

Real-time operations platform augmenting ERP, CRM, and MES with production scheduling, machine utilization, scrap analytics, quality tracking, and delivery predictions.

Intelligence layer above existing systems providing the visibility executives need without replacing what works.

EPC projects are document-heavy and fragmented. Engineers spend hours on calculations requiring manual data extraction from PDFs and vendor documents.

50+ engineering calculation and analytics tools across nine domains, combining deterministic calculations with AI-assisted document extraction and vendor deviation analysis.

Entire 50+ tool suite built in a single sprint using a parallelized development approach.

A bank's mortgage and student loan approval process required manual review. High-quality applicants faced the same slow process as complex cases.

ML models that fast-track high-quality applicants with fewer reviewers while routing complex cases to senior staff. Governance and audit trails built from day one.

Shipped ML models in a regulated banking environment by building governance, documentation, and audit trails from the start.

A bank evaluated each product independently. Some accounts appeared unprofitable in isolation, but account holders brought wealth management, loans, and referrals.

Relationship-level analytics evaluating the entire client relationship across all products, accounts, and referral chains using graph representations.

Transformed decision-making from product-level to relationship-level portfolio thinking.

Regulated banks must submit stress test results demonstrating capital adequacy under adverse scenarios, under direct regulatory scrutiny.

Quantitative models for scenario analysis, loss estimation across loan portfolios, revenue projection under stress, and capital adequacy analysis meeting regulatory standards.

Operated under direct FDIC scrutiny where every model decision must be documented, validated, and defensible.

Corporates faced complex, multi-dimensional risk exposures across currencies, commodities, interest rates, and equity that standard products could not address.

Bespoke structured solutions pairing simulation engines with hedging strategies and capital-efficient vehicles, custom to each client's risk profile.

Translated complex multi-asset correlated risk into products that C-suite executives could understand. $10M+ in direct value.

Corporate treasurers and CFOs needed to understand the impact of market movements on balance sheets, pension obligations, and hedging positions.

Balance-sheet simulation and scenario analysis engines translating complex models into board-level decision tools for capital allocation and M&A valuation.

Quantitative models made accessible as board-level decision tools shaping billion-dollar capital allocation.

Corporate treasury functions needed to optimize cash management, liquidity positioning, and investment of surplus funds across global operations.

Quantitative models for cash flow forecasting, liquidity buffer sizing, investment portfolio construction, and FX exposure management across subsidiaries.

End-to-end treasury optimization from cash forecasting through investment and FX management.

A company needed a clearly articulated AI transformation strategy aligned with the board, covering prioritization, governance, and investment.

Enterprise AI strategy presented to the board covering portfolio prioritization, governance framework, investment criteria, and measurable KPIs.

Translated AI capabilities and limitations into language the board could use for capital allocation and risk management.

A large distributor growing through acquisitions needed to unify AI and data strategy for rapid integration of acquired companies.

Enterprise AI and data vision centralizing governance and integrating supply chain, pricing, and sales analytics with standardized integration playbooks.

Aligned M&A strategy with AI capabilities, creating repeatable integration playbooks.

An enterprise lacked centralized experimentation. Teams ran tests without statistical rigor, leading to false conclusions.

Experimentation CoE responsible for enterprise-wide A/B testing methodology, experiment design, statistical analysis, and incrementality measurement.

Replaced ad-hoc testing with rigorous, enterprise-wide experimentation standards.

Large enterprises suffer from siloed data where requests get diluted through management layers, producing locally optimal but globally suboptimal decisions.

Integrated decision platform connecting data across business functions for unified, enterprise-wide decision-making.

Cross-functional data integration replacing siloed optimization with enterprise-wide decision support.

A bank needed to scale its premium client service model. Analytics insights were not reaching the people making daily decisions.

Embedded cross-functional data teams in every major business unit with decision tools tailored to each function's workflows.

Embedded data people directly in business functions so insights became part of daily decisions, not periodic reports.

Traditional software development is serial and slow. AI agents can plan, build, review, and ship in parallel, but need governance for quality.

Parallelized pipeline where an AI architect produces specifications, multiple AI workers build in parallel, and an AI reviewer validates before integration.

50+ production-quality tools delivered in a single sprint. Engineering throughput multiplied without sacrificing quality.

Small and medium-sized financial institutions lack access to modern core banking systems that handle their specific operational and regulatory needs.

Core banking platform with double-entry ledger, loan management, member shares, fixed deposits, regulatory reporting, and full accounting suite.

Modern core banking built for underserved small and medium financial institutions.

Semiconductor companies needed mathematical proof that chip designs were correct before committing to fabrication, where a single bug costs millions.

Verification tools and algorithms for proving correctness of chip designs. 8 granted patents.

A decade of deep work in formal methods, mathematical logic, and verification. Demonstrates the ability to solve hard mathematical problems in industrial settings.