Associate Partner, McKinsey & Co · Senior Principal, QuantumBlack

Turning Complex AI
Ecosystems into Production Outcomes.

Senior technology leader trusted to architect, deliver, and scale enterprise AI systems in demanding environments.

Capabilities

What I Solve

The hard problems. The ones that stall when complexity meets ambition.

01
AI from Pilot to Production
Most enterprises can run a pilot. Few can scale it. Designing the architecture, governance, and organizational capability that takes AI from controlled experiment to governed production system—with measurable outcomes on the other side.
02
Complex Ecosystem Delivery
Enterprise AI rarely fails on the model. It fails on the ecosystem: misaligned stakeholders, legacy systems, governance gaps, and competing priorities. Leading delivery where complexity is the environment, not the exception.
03
Enterprise Architecture & Platforms
Designing scalable data platforms, AI workflows, agentic systems, and operating models that support production AI today and adapt to what's coming. Architecture decisions made with delivery, not just elegance, in mind.
04
Decision Intelligence Systems
Forecasting, optimization, anomaly detection, and digital twins—connecting advanced analytical capability, both traditional ML and GenAI, to the decisions that actually move the business. Systems built for operational use, not executive dashboards.
Delivery Profile

Trusted in Complex Environments

◆ Leadership
Cross-functional from C-suite to engineering
Aligning executives, architects, data teams, risk, and operations toward shared delivery outcomes.
◆ Delivery
Production systems, not prototypes
A decade at McKinsey & QuantumBlack focused on getting AI from controlled environments into governed, monitored, production-grade deployment.
◆ Breadth
Global enterprise delivery experience
Complex programs across geographies, regulatory environments, and organizational structures.
◆ Technical
Architecture judgment, not just oversight
Deep technical fluency across traditional ML, GenAI, agentic systems, data platforms, and enterprise architecture.
◆ Value
ROI-driven transformation focus
Every engagement oriented toward measurable business outcomes—cost, speed, revenue, decision quality. Not experimentation for its own sake.
◆ Execution
Bridges strategy, architecture, and delivery
Rare combination of executive alignment, technical depth, and hands-on delivery discipline in a single operator.
20+
Years across data, analytics, and AI
A Decade
Of McKinsey & QuantumBlack leadership
Global
Enterprise-scale programs across industries and geographies
Hands-On
Still building and experimenting with frontier systems
Domain Focus

Active Capabilities

A focused set of areas where deep expertise meets sustained investment.

Agentic AI Systems
Enterprise AI Operating Models
Memory Architectures
Scalable GenAI Platforms
Decision Intelligence
Optimization Engines
Digital Twins
AI Governance at Scale
Semantic Layers & Knowledge Systems
Forecasting & Anomaly Detection
Responsible AI
Data Platform Architecture
Career

Experience

Two decades of progressive leadership across analytics, data, and AI delivery in large, complex organizations.

Current
McKinsey & Company
Associate Partner
Current
QuantumBlack, AI by McKinsey
Senior Principal
Prior
Metro / Makro Cash & Carry
Manager, Data Mining
Prior
dunnhumby
Data Solutions Manager
Prior
Target
Analyst
Prior
HSBC
Business Analyst
Operator Intelligence

Featured Insights

Data Platforms New
The Enterprise Knowledge Stack
Taxonomy, ontology, semantic mapping, and knowledge graph are four layers of one architecture—not four names for the same thing. The complete technical walkthrough, with diagrams for each layer.
Read article
AI Strategy Featured
Which Model Wins Is Not the Question. Who Controls the System Is.
Frontier model access is not a utility—it is a dependency. Most enterprise AI programs have no architecture for when that dependency goes dark overnight.
Read article
Delivery 7 min read
Why Most Enterprise AI Programs Fail After the Pilot
The demo worked. The pilot delivered. And then momentum died. This failure pattern follows a consistent, diagnosable script—and it's not a technology problem.
Read article
Let's Connect
Always Happy to Connect

Whether to discuss enterprise AI, exchange perspectives on what's working in production, or just compare notes — always happy to have a substantive conversation.

Delivery Portfolio

Where Complex Technology
Meets Real Delivery

Two decades of enterprise delivery across analytics, AI transformation, data platforms, decision systems, and operating model design. Engagement details are confidential; the nature of the work is not.

01
Enterprise AI Transformation
Turning fragmented AI initiatives into coordinated, scaled operating capability. From governance design and sequencing to delivery execution—building the foundation that makes AI programs durable rather than episodic. Includes standing up AI COEs, prioritization frameworks, and production deployment programs that span multiple business units.
Strategy Governance Execution COE Design
02
Production GenAI Systems
Moving generative AI from controlled experiments into governed, monitored production workflows. Designing the architecture, testing frameworks, and oversight mechanisms that make GenAI trustworthy at scale. Includes agentic system design, RAG pipelines, memory architectures, and integration with enterprise data and process layers.
GenAI Agentic Systems RAG Production
03
Data Foundations for AI
Building the data platforms, semantic layers, and governance structures that make AI reliable in production. Addressing the gap between analytics-ready and genuinely AI-ready data infrastructure. Includes data product design, lakehouse architecture, data quality frameworks, and lineage systems that support governed AI deployment.
Data Platform Semantic Layer Governance Lakehouse
04
Decision Intelligence
Forecasting, anomaly detection, optimization, and simulation—connecting advanced analytical capability to the decisions that move the business. Designing systems that make decision-making faster, more consistent, and measurable. Includes both traditional ML and GenAI approaches, and hybrid architectures where each plays to its strengths.
Forecasting Optimization Digital Twins Traditional ML
05
AI Operating Models
Designing the organizational structures, governance frameworks, and ways of working that allow AI to scale beyond isolated teams. COE design and implementation, hub-and-spoke talent models, prioritization and portfolio management, and value measurement systems that sustain executive commitment to AI investment over time.
Operating Model COE Talent Architecture Value Tracking
06
Transformation Rescue & Scale-Up
Helping stalled or struggling AI programs regain momentum. Diagnosing what's actually broken—governance, ownership, architecture, change management—and building a recovery path that creates durable progress rather than renewed activity. Includes program assessment, stakeholder realignment, and execution recovery.
Program Recovery Diagnosis Stakeholder Alignment Execution
Approach

How the Work Gets Done

Delivery discipline applied to technically complex, organizationally demanding environments.

01 · Diagnosis
From clarity to momentum
Every engagement starts with a clear diagnosis of what's actually constraining progress—not the presenting problem, but the structural root cause. Technical problems get technical solutions. Organizational problems get organizational intervention. The distinction matters.
02 · Orchestration
Cross-functional delivery leadership
Enterprise AI delivery requires aligning functions that don't naturally align: business, engineering, data, risk, legal, and operations. The leadership capability that makes this work—stakeholder alignment, sequencing, change management—is as important as the technical work.
03 · Production
Outcomes over activity
Every engagement is oriented toward measurable outcomes in production: systems that run, decisions that improve, organizations that are more capable at the end than the beginning. Not demos. Not strategy decks. Not pilots that never graduate.
Always happy to connect
If you're working on something in this space and want to exchange perspectives, compare notes, or just have a substantive conversation — reach out.
Operator Intelligence

Perspectives from the
Delivery Side of Enterprise AI

What actually works in production—and why most programs struggle to get there. Earned perspectives, not commentary.

Data Platforms July 2026
The Enterprise Knowledge Stack: Taxonomy, Ontology, Semantic Mapping, Knowledge Graph
Four different layers of one architecture, not four names for the same thing. A complete technical walkthrough—what each layer is, how it is represented, how they connect, and the order to build them—with diagrams for each.
13 min read Read →
AI Strategy June 2026
The Next Enterprise AI Question Is Not Which Model Wins. It Is Who Controls the System.
On June 12th, a government directive disabled Fable 5 globally at 5:21pm on a Friday. Frontier model access is not a utility. It is a dependency. Most enterprise AI programs have no architecture for when that dependency goes dark overnight.
7 min read Read →
Agentic AI June 2026
MCP and A2A Are Building the TCP/IP Moment for Enterprise AI
Before TCP/IP, every network connection was a custom integration. N networks required N×(N−1) custom integrations. The same problem is now playing out in enterprise AI. MCP and A2A are solving it at the protocol level — and the adoption numbers suggest the enterprise world is recognizing it.
6 min read Read →
AI Strategy June 2026
The $5 vs $0.19 Problem: Why Open Models Are Rewriting Enterprise AI Economics
GPT-4o costs $5.00 per million tokens blended. DeepSeek V3 costs $0.19. That is a 26x cost difference for a 1.6 percentage point capability gap. Every enterprise AI program should be running a tiered model portfolio. The ones that are not will face an uncomfortable conversation with the CFO.
5 min read Read →
Architecture June 2026
World Models and the Digital Twin: When AI Learns to Simulate Reality
AI systems that learn to predict and simulate how the world evolves are converging with industrial digital twins. BMW's Virtual Factory is projected to reduce production planning costs by 30%. This is a production deployment across a global manufacturing network — not a research project.
5 min read Read →
Architecture June 2026
Model Routing Is Not a Trick. It Is an Architecture Decision.
The RouteLLM paper quantifies 3.66x cost reduction at 95% of frontier quality. At half that — 1.8x — across an AI program spending $1M per year on inference, that is $550,000 in annual savings from an architecture change that takes 30 days to implement.
6 min read Read →
AI Strategy May 2026
The Model Is Not the Moat. The Harness Is.
The performance gap between frontier models has compressed to the point where the underlying model is increasingly a commodity. Durable advantage lives in proprietary data, domain-specific context, embedded workflows, and evaluation feedback loops. The harness is the business.
5 min read Read →
Data Platforms May 2026
Synthetic Data Is Not a Shortcut. It Is an Infrastructure Decision.
Microsoft's Phi-4 states it directly: synthetic data has direct advantages over organic data. A 14B model trained with synthetic data surpasses GPT-4 on STEM benchmarks. The argument is not that synthetic data is cheaper. It is that it can be engineered to be better.
5 min read Read →
Architecture May 2026
Diffusion Models Are Escaping the Image Lab. The Enterprise Implications Are Significant.
The same iterative denoising mechanism behind image generators is now producing picomolar-affinity protein binders, physically coherent robot trajectories, and combinatorial optimization solutions. RFdiffusion reduced candidate screening from tens of thousands to as few as one. This is production-accessible today.
5 min read Read →
Delivery April 2026
Why Most Enterprise AI Programs Fail After the Pilot
The demo worked. The pilot delivered. And then momentum died. This failure pattern is more common than the industry admits—and it follows a consistent, diagnosable script that has nothing to do with the model.
7 min read Read →
Agentic AI April 2026
The Case for the Single Orchestrating Agent
Most enterprise agent programs are over-architected before they've proven anything. One strong, well-instrumented agent that reliably executes multi-step work is worth more than a network of agents that can't be trusted.
7 min read Read →
Architecture April 2026
What Production-Grade Agent Systems Actually Require
The demo worked. Everyone agreed the technology was ready. Then the production question came up—and the answers got vague. Reliability, observability, governability, and recoverability are not optional.
6 min read Read →
Delivery January 2026
How to Rescue a Stalled AI Program
Stalled AI programs are more common than the industry admits. The pattern is recoverable—but recovery requires honest diagnosis, not more technology investment. Most diagnoses are looking in the wrong place.
6 min read Read →
Data Platforms November 2025
The Semantic Layer Is the Missing Infrastructure in Most Enterprise AI Programs
MCP connects your agents to your tools. But without consistent semantic definitions underneath, those connections produce inconsistent outputs. The semantic layer is the piece most programs are missing.
7 min read Read →
Agentic AI May 2025
How to Evaluate an AI Agent Before You Trust It With Real Work
Accuracy on a test set is not a production evaluation. Agents operate in environments with edge cases, adversarial inputs, and cascading decisions. Here is how to actually evaluate one.
6 min read Read →
Strategy January 2025
DeepSeek Changed the Conversation About AI Cost — Now What?
DeepSeek R1 training reportedly cost $6 million versus hundreds of millions for comparable US models. The cost narrative around AI has changed permanently. Here is what that means for enterprise strategy.
5 min read Read →
Architecture December 2024
The Open vs Closed Model Question Has Changed — Here Is the New Calculus
In 2023, the question was capability. Open models were behind. In 2024, the capability gap closed for most enterprise use cases. Now the question is about something else entirely.
6 min read Read →
Agentic AI November 2024
What Anthropic's MCP Actually Means for Enterprise AI Architecture
MCP is not a product. It is a protocol. And protocols that win create ecosystems. Here is why MCP matters for how enterprises should think about their AI integration architecture.
5 min read Read →
Data Platforms March 2024
RAG Is Not a Data Strategy — It Is a Symptom of One
RAG architectures became ubiquitous in 2024. But a RAG system that retrieves from a poorly governed data store is retrieving the wrong things with high confidence. The data problem does not go away.
5 min read Read →
Delivery October 2023
Taking LLM Applications to Production: Latency, Cost, and Evaluation Engineering
Getting an LLM demo working takes a weekend. Getting it to production takes engineering around token economics, latency budgets, evaluation datasets, and output validation.
7 min read Read →
Architecture May 2023
Retrieval-Augmented Generation: The Architecture Behind Grounded LLM Applications
RAG grounds LLM outputs in your own data: chunking, embeddings, vector search, prompt assembly, and the evaluation metrics for each stage. A technical walkthrough of the full pipeline and where it fails.
7 min read Read →
Architecture September 2022
Transformers and Transfer Learning in the Enterprise: From BERT to GPT-3
The pretrain-and-finetune paradigm changes the economics of enterprise NLP: what encoder and decoder models are each good at, how to fine-tune with small labeled sets, and where classical ML still wins.
7 min read Read →
Data Platforms February 2022
Data Mesh: What Domain-Oriented Data Architecture Actually Requires
Data mesh is an organizational architecture as much as a technical one: domain ownership, data as a product, self-serve platform, federated governance — and the prerequisites without which it fails.
7 min read Read →
Data Platforms October 2021
Feature Stores: Solving Training-Serving Skew and Point-in-Time Correctness
Two of the hardest ML infrastructure problems — features computed differently in training and serving, and future information leaking into training data — have a shared architectural answer.
7 min read Read →
Delivery March 2021
Why Machine Learning Models Degrade in Production: Data Drift, Concept Drift, and What to Monitor
A model is a snapshot of the relationship between inputs and outcomes at training time. When the world moves, that snapshot goes stale. The drift taxonomy, detection methods, and retraining strategies.
7 min read Read →
Architecture December 2020
Understanding the Vanishing Gradient Problem (VGP) and Solutions
VGP makes it difficult to train the parameters of early layers in deep neural networks. This article breaks down the cause — saturating activation functions — and the solutions: ReLU, LSTM, GRU, and alternative weight matrices.
3 min read Read →
Architecture December 2020
Deep Learning Frameworks: Which One Is Right for Your Problem?
TensorFlow, PyTorch, Keras, Caffe, MXNet, CNTK — each framework was built for specific objectives. A practical guide to what each one does well and how to choose the right one for your problem statement.
6 min read Read →
Delivery May 2020
Five Steps for Successful AIOps Adoption and Scaling
30% of large corporations are projected to use AIOps exclusively by 2023. Most organizations struggle to move AI projects into production. A five-step framework — from problem scoping through sustained production management — for AIOps use cases that actually deliver business impact.
4 min read Read →
Data Platforms September 2019
Using Feedback Loops to Correct Labels and Acquire Ground Truth
Supervised ML depends on labeled datasets, but labeling is error-prone. A feedback loop framework — where a model's predicted outputs are used to correct labels and acquire ground truth — is one of the most practical tools for improving data quality over time.
4 min read Read →
Data Platforms July 2019
Inferential Statistics in Python: A Practical Reference Guide
Descriptive statistics summarize data. Inferential statistics let you make hypotheses about a sample and apply them to a population. Covers probability distributions, hypothesis testing, Type 1 and Type 2 errors, confidence intervals, and correlation — with Python implementations throughout.
7 min read Read →
Data Platforms January 2019
Data Leakage in Machine Learning: What It Is, Why It Happens, and How to Avoid It
A model that performs suspiciously well in development and then fails completely in production is the classic symptom of data leakage. Covers both types of leakage, how to detect it through EDA and feature importance, and how to prevent it at every stage of the ML lifecycle.
5 min read Read →
Architecture January 2019
Vector and Matrix Mathematics for Machine Learning in Python
Mathematics is a basic requirement for machine learning — especially deep learning. This post covers scalars, vectors, matrices, and core NumPy operations as a practical reference for the linear algebra that underpins ML algorithms.
3 min read Read →
Architecture December 2018
ARIMA vs LSTM for Time Series Forecasting: When to Use Which
Traditional ARIMA methods forecast time series data with high accuracy. LSTM offers superior performance at the cost of increased complexity. A practical guide to both approaches — with implementation in R (ARIMA) and Python (LSTM) — and when to choose each.
7 min read Read →
Data Platforms June 2016
Class Imbalance in Machine Learning: Techniques and What Actually Works
Everyone says to balance your classes before training a classifier. But how much does it actually help? An experiment with logistic regression, random forest, and SVM — with results showing that balancing is not a guaranteed improvement and the benefit varies significantly by algorithm.
4 min read Read →
Architecture May 2016
Regularization in Machine Learning: L1, L2, and Elastic Net with R and SAS
Regularization prevents overfitting by penalizing model complexity. Ridge (L2) shrinks coefficients. Lasso (L1) shrinks and selects features. Elastic Net combines both. A practical guide to the concept and implementation in R (glmnet) and SAS (PROC GLMSELECT).
5 min read Read →
Technical Journal

The Build Journal

Selective experiments, prototypes, and system notes from staying close to the frontier. Not polished thought leadership—raw signal from hands-on work with emerging technology.

July 2026 Semantic Canvas
Ontology · Applied Build
Semantic Canvas — turning the four-layer framework into a working knowledge designer
The knowledge stack article kept generating the same follow-up: the layers make sense, but what does authoring them actually feel like? So I built the tool. Semantic Canvas takes a plain-English domain description and drafts the semantic foundation — domains, glossary terms, entity types, relationships, attributes — as proposals in a review queue, never as accepted truth. One canonical model underneath; the taxonomy tree, business glossary, ontology canvas, and property-graph schema are four projections of it — the article's core argument, enforced in code. Two design decisions carried the build: every concept keeps provenance (user input vs. AI generation vs. import vs. manual edit), and a deterministic validation engine runs continuously — circular hierarchies, near-duplicate terms, entities without identifiers, orphaned concepts. The most useful experiment was import: feed it a legacy CSV glossary and it detects which rows collide with the model, shows the conflicting definitions side by side, and offers a merge that preserves both wordings for the steward. That single workflow — load a real glossary, surface the conflicting definitions in minutes — lands harder in conversations than any diagram. Runs entirely in the browser; the demo needs no account and stores nothing server-side.
July 2026 Knowledge Stack
Ontology · GraphRAG
Building the four-layer knowledge stack end to end — and testing GraphRAG against vector-only retrieval
The ontology drift work from March kept pulling me back to a more basic question: what do the layers underneath a knowledge graph actually need to look like? Built a working miniature of the full stack on a mock product domain — a SKOS-style taxonomy, extended into a small ontology with classes, typed relations, and constraints; two mock source schemas mapped against it; and a property graph populated through those mappings with constraint validation at load time. Then the test that mattered: multi-hop questions against GraphRAG-style retrieval (traverse typed edges, hand the subgraph to the model) versus flat vector retrieval over the same corpus. Vector-only holds up on single-fact lookups and fails hard on relational questions — anything shaped like "which X are affected by Y through Z" — where graph traversal answers deterministically. The underrated win was validation: rejecting nonconforming instances at load caught modeling errors that would otherwise have surfaced as mysterious retrieval misses weeks later. Wrote the full framework up as an article.
April 2026 Memory Systems
Agentic AI
Testing memory patterns across modern agent frameworks
Comparing episodic, semantic, and procedural memory patterns across major agent frameworks. The core tension: how much context to persist vs. how much to re-derive at runtime. More persistence means better continuity but higher latency and cost; more re-derivation means cleaner state but agents that feel amnesiac after short gaps. Building a lightweight evaluation harness to compare approaches on realistic multi-session enterprise tasks. Early signal: most frameworks optimize for single-session performance and treat cross-session continuity as an afterthought.
April 2026 Knowledge Systems
RAG · Knowledge Graphs
Building private AI wiki systems from unstructured folders and notes
End-to-end pipeline from raw unstructured documents—PDFs, markdown notes, email threads, meeting transcripts—into a queryable, agent-accessible private knowledge system. Exploring chunking strategies (fixed-size vs. semantic), embedding model choices, and hybrid retrieval (dense + sparse) for precision on domain-specific queries. The hard part isn't ingestion—it's freshness. Building an incremental re-indexing layer that detects when source documents change and updates only affected chunks, without full re-ingestion. Also experimenting with graph-based knowledge structures alongside vector search to handle entity relationships that flat retrieval misses.
March 2026 Optimization
Traditional ML · Decision Intelligence
Comparing optimization-first scheduling engines vs. LLM planning approaches
Running a head-to-head between constraint-based solvers and language model planners on a realistic resource scheduling problem with mixed hard and soft constraints. LLMs are surprisingly competitive on soft constraint handling and produce far better natural-language explanations. But they degrade badly as hard constraints multiply—they start forgetting constraints mid-plan in ways that are hard to detect without systematic validation. Hybrid architecture looks most promising: LP/MIP solver handles hard constraint satisfaction, LLM handles user-facing explanation and soft constraint ranking. Neither alone is the answer.
March 2026 Ontologies
Knowledge Engineering · Graph Memory
Exploring enterprise ontology evolution and graph memory patterns
How do you keep a knowledge graph current as business domains evolve? Products get renamed. Org structures change. Definitions drift. Experimenting with LLM-assisted ontology curation—using models to detect when incoming documents introduce terminology that conflicts with or extends the existing graph. Also testing automated drift detection: flagging nodes whose connected context has shifted significantly over a rolling time window. Working on a lightweight proposal-and-approval workflow for ontology changes that doesn't require dedicated knowledge engineers.
February 2026 Toolchains
Local AI · Agent Workflows
Hands-on experiments with local AI toolchains and agent workflows
Running capable quantized models locally for private, low-latency agent tasks. Evaluating practical trade-offs against cloud APIs across four dimensions: privacy (local wins clearly), cost at scale (local wins after break-even), latency (local wins for short context), and capability ceiling (cloud still ahead, gap narrowing). Tool-use reliability is the biggest gap—smaller local models still struggle with consistent structured output under complex chaining. Careful prompt engineering and explicit state management compensate for a surprising amount of raw capability difference.
February 2026 Observability
Enterprise AI · Governance
Designing enterprise-grade agent observability and control layers
What does production-grade observability look like for autonomous agents running consequential enterprise tasks? Standard software observability doesn't translate cleanly to non-deterministic systems. Building a prototype that captures structured decision traces, tool call logs with full parameter and response capture, escalation records, and anomaly flags. Key design challenge: making traces human-readable without losing fidelity. Using a two-layer approach—structured JSON trace for machine processing, LLM-generated plain-English summary for human auditors. Next step: integrating with a human-in-the-loop approval workflow for flagged decisions.
January 2026 Hybrid ML
Traditional ML · GenAI Integration
Prototyping hybrid ML + GenAI pipelines for structured prediction tasks
Testing a pattern that's becoming more relevant as GenAI matures: using traditional ML models as guardrails and validators for GenAI outputs, rather than treating them as alternatives. A gradient boosting model trained on historical outcomes flags GenAI-generated recommendations that fall outside the distribution of historically valid decisions. The combination catches hallucinations and out-of-distribution outputs more reliably than using a second language model as a checker—and is more explainable to business stakeholders. Also experimenting with using GenAI to generate features and structured context that improve traditional ML performance on sparse or unstructured input domains.
August 2025 AIOps Revisited
Traditional ML · LLMs
Revisiting my 2019 incident-management ML work with LLM approaches
Went back to the incident event analysis I built in 2019 — classification and assignment use cases over incident event logs — and re-ran the problem with current tools. Zero-shot and few-shot LLM classification against the trained classifiers from the original pipeline: the LLM wins on cold start and the long tail of rare categories with no labeled data at all, while the classical models remain far cheaper, faster, and better calibrated on the high-volume head. The practical answer is the split — classifier for the head, LLM for the tail, plus resolution-summary drafting the old pipeline could never do. Six years between two approaches to the same dataset is a useful measure of how much has changed, and how much has not: data quality and label taxonomy were the constraint in 2019, and they still are.
September 2024 Local Models
Local AI · RAG
First serious pass at local quantized models and a personal RAG prototype
Spent evenings running quantized 7-8B models locally through Ollama — Llama 3 8B and Mistral 7B at 4-bit — to get a firsthand feel for what fits on a laptop and what breaks. Tokens per second are fine for chat use; structured output reliability degrades noticeably under quantization. Also built a first RAG prototype over my own notes archive: naive fixed-size chunking, local embeddings, cosine retrieval. The lesson everyone reports but you only internalize by building it yourself: generation quality was never the problem — retrieval quality was, and chunk boundaries drove most of the failures. This small prototype is where the later knowledge-systems work started.
Public & Advisory

Speaking on the Future
of Enterprise AI

Available for select conferences, podcasts, panels, leadership forums, and private executive sessions where the conversation is substantive.

Topics

What I Speak On

All topics grounded in real enterprise delivery—not theoretical frameworks or vendor talking points.

01
Agentic AI in the Enterprise: What's Real, What's Not
02
How Fortune 500 Firms Are Actually Adopting AI
03
AI Beyond Pilots: What Actually Scales
04
Memory Systems and the Next Wave of AI Agents
05
AI Operating Models for Large Enterprises
06
Decision Intelligence & Digital Twins in Practice
07
Separating AI Hype from Durable Business Value
08
Enterprise Architecture for the Agent Era
Formats
Keynotes Panel discussions Fireside chats Podcast conversations Executive offsites Leadership briefings Advisory sessions Board presentations
ST
Sanchit Tiwari
Associate Partner, McKinsey & Company · Senior Principal, QuantumBlack
Sanchit Tiwari is a senior enterprise technology delivery leader with two decades of experience helping large organizations turn complex AI programs into production outcomes. He serves as Associate Partner at McKinsey & Company and Senior Principal at QuantumBlack, AI by McKinsey.

His work spans generative AI, agentic systems, decision intelligence, digital twins, and the operating models required to take AI from isolated pilots to governed, production-grade programs at scale.

Known for bridging technical depth with executive communication, Sanchit brings a practitioner's perspective to every conversation—grounded in hands-on delivery across industries and geographies.
Request a Speaking Inquiry

Share brief details and Sanchit will follow up within a few business days.

About

Sanchit Tiwari

Sanchit Tiwari works where AI ambition meets execution reality. He helps large organizations move from pilots and fragmented initiatives to scalable production systems with measurable business value.

He combines executive alignment, architecture judgment, and delivery discipline—the combination that determines whether AI programs create lasting capability or expensive artifacts.

He has led enterprise-scale programs spanning generative AI, agentic systems, traditional ML, advanced analytics, digital twins, decision intelligence, scalable data platforms, and AI operating models. His work is cross-functional by nature: aligning business leaders, engineering teams, data organizations, risk functions, and operations toward shared production outcomes.

Known for translating complexity into momentum, Sanchit operates at the intersection of strategy and execution. He is as comfortable in an architecture review as in an executive steering committee—and understands that the two conversations have to connect for programs to succeed.

He maintains hands-on technical fluency—actively experimenting with frontier systems, prototyping architectures, and staying close to what's actually emerging—so that his strategic and delivery judgments are grounded in current technical reality, not abstracted from it.

Quick Profile
◆
Current Role
Associate Partner
McKinsey & Company
◆
Also
Senior Principal
QuantumBlack, AI by McKinsey
◆
Location
Greater Chicago Area
United States
◆
Focus
Enterprise AI Delivery
Production Systems · Architecture · Operating Models
◆
Experience
20+ Years
Analytics, Data & AI Delivery Leadership
Expertise

Areas of Deep Work

Two decades of progressive investment across a focused set of interconnected domains.

Enterprise AI Delivery
End-to-end leadership of large-scale AI programs—from strategy and architecture through governed production deployment and adoption.
Agentic & Generative AI
Architecture and delivery of agentic systems, LLM-based workflows, RAG pipelines, and GenAI platforms for enterprise production use cases.
Traditional ML & Statistical Modeling
Deep experience with regression, classification, time series forecasting, clustering, and optimization—the methods that still power high-stakes production decisions.
ML + GenAI Integration
Designing hybrid architectures where traditional ML and generative AI each play to their strengths—combining precision, explainability, and flexibility in the same system.
Decision Intelligence
Combining ML, optimization, and business logic to create systems that support faster, more consistent, and more autonomous decisions at scale.
Data Platform Architecture
Scalable, AI-ready data infrastructure: lakehouses, semantic layers, data products, lineage systems, and real-time pipelines designed for production AI.
AI Operating Models
Organizational structures, governance frameworks, and ways of working that enable AI to scale beyond isolated programs into enterprise-wide capability.
Digital Twins & Simulation
Digital twin architectures for operational planning, scenario analysis, continuous optimization, and decision support in complex operational environments.
AI Governance & Responsible AI
Risk frameworks, model risk management, explainability standards, audit architectures, and compliance design for production AI in regulated environments.
Contact

Let's Have a
Real Conversation

If you're working on enterprise AI and want to exchange perspectives, discuss what's actually working, or just connect — feel free to reach out. All messages read personally.

in
LinkedIn
Connect professionally
→
🛠
GitHub
Experiments and build projects
→
◆
Delivery Leadership
Enterprise AI programs and strategy
→
Location
Greater Chicago Area, United States
Send a Message

Use the form below for direct inquiries. Include context on what you're working on or trying to solve.

✓  Message received. Sanchit will be in touch within a few business days.