Sumanta Mukhopadhyay · Fractional CTO & AI ArchitectFrom the ontology to the infrastructure it runs on.
I'm a principal engineer and fractional CTO. I help leadership teams turn AI and data ambitions into systems that actually ship: the domain model, the architecture, the platform and the team that runs it.
Technical leadership without the gap between slides and code.
Most organisations have strategists who can't build and builders who aren't in the room when strategy is set. I work across both, so the decisions that get made are ones the system can actually honour.
Technology strategy & fractional CTO
Own the technical agenda at board and executive level, and make it real in the engineering org.
- Roadmaps tied to business outcomes
- Build vs. buy vs. partner decisions
- Hiring, team topology, engineering culture
- Technical due diligence
Applied AI & agentic systems
Move from demos to dependable AI products with evaluation, guardrails and cost control built in.
- RAG and GraphRAG pipelines
- LangChain / LangGraph multi-agent workflows
- MCP tool integration & agent ops
- LLM evaluation, observability, governance
Knowledge & semantic architecture
Give your data a shared meaning so humans, analytics and AI agents all reason over the same model.
- Ontology & taxonomy design
- Knowledge graphs, entity resolution
- Standard vocabularies & common data models
- Semantic layers for analytics and LLMs
Data platforms at scale
Design the pipelines and stores that carry the business. Real-time and batch, reliable under load.
- Event-driven & streaming architectures
- MongoDB, Postgres, BigQuery, warehouses
- Data modelling & aggregation design
- Quality, lineage and governance
Cloud & platform engineering
Infrastructure that is reproducible, secure and pleasantly boring, on any cloud.
- Kubernetes & internal developer platforms
- Terraform IaC across AWS, Azure, GCP
- CI/CD, GitOps, observability, SRE
- Cost and reliability engineering
Product engineering
Hands-on delivery of the product surface: APIs, web and mobile, built to the architecture's standard.
- Python, Go, TypeScript, Java, C#
- React, Angular, React Native
- API design & system design
- Developer tooling & CLIs
The problem you have, and how I'd solve it.
Each engagement follows the same disciplined method. It adapts to the problem, but you always know what happens next, what it costs and what you get.
Discover
Stakeholder interviews, system and data audit, constraints.
Model
Map the domain into an ontology everyone agrees on.
Architect
Target architecture, decision records, phased roadmap.
Build
A thin production slice first, then expand with evidence.
Scale
Load, cost, reliability and security hardened to target.
Transfer
Runbooks, hiring and mentoring until the team owns it.
Our AI pilot won't make it to production
Demos impress, but answers are unreliable, costs are unknown and nobody will sign off.
Usually the model isn't the problem. Retrieval is ungrounded, there's no evaluation set, and no one owns quality.
- Define the decisions the AI supports and the cost of a wrong answer
- Ground retrieval in a domain ontology and knowledge graph (GraphRAG)
- Build an evaluation harness and tracing before adding features
- Orchestrate agents with LangGraph: explicit state, tools via MCP, human checkpoints
- Ship one workflow to production behind guardrails, then expand
Our data is everywhere and means different things
Every team has its own definition of a customer, a patient or revenue. Reports disagree. AI makes it worse.
It's a semantic problem, not a storage problem. Without a shared model, more pipelines just copy the confusion.
- Run domain workshops to capture core concepts and relationships
- Formalise an ontology and controlled vocabularies, versioned like code
- Map sources onto it: entity resolution and master data
- Expose a semantic layer to BI, APIs and AI agents
- Put data contracts and quality checks at the boundaries
Our platform buckles as we grow
Latency spikes at peak, batch jobs overrun, incidents are frequent and every release feels risky.
Synchronous coupling and shared databases turn growth into contention. The architecture is fighting the load.
- Instrument first: find the real bottlenecks with tracing and load tests
- Set SLOs and latency budgets for the critical paths
- Move hot paths to an event-driven core with idempotent consumers
- Redesign data access: indexes, aggregations, read models, caching
- Automate deploys and rollbacks on Kubernetes with GitOps
Our healthcare data won't interoperate
EHRs, labs and claims arrive in different formats and codes. Every integration is a bespoke project.
Point-to-point integration doesn't scale. You need a standard exchange format and a standard analytical model.
- Inventory sources, code systems and PHI flows
- Standardise exchange on HL7 FHIR; normalise to SNOMED, LOINC and RxNorm
- Land analytics in OMOP CDM for reproducible research with OHDSI tools
- Build in de-identification, consent, encryption and audit trails
- Layer clinical AI only on governed, provenance-tracked data
Cloud costs and complexity are out of control
Environments were hand-built, nobody knows what's running, and the bill grows faster than revenue.
Without infrastructure as code and cost visibility, every environment drifts and waste is invisible.
- Map every resource to an owner, a service and a cost
- Codify infrastructure in Terraform modules across AWS, Azure and GCP
- Consolidate workloads onto a shared Kubernetes platform
- Right-size, autoscale and route AI calls to the cheapest capable model
- Report unit economics per transaction, tenant and token
We don't have senior technical leadership
The founders or executives are making technical bets without a trusted, experienced voice in the room.
You don't need a full-time CTO yet. You need CTO-grade judgement on the decisions that are expensive to reverse.
- Assess the team, architecture, roadmap and risk in the first month
- Set a technical strategy tied to business milestones
- Establish architecture governance: ADRs, reviews, standards
- Hire and mentor senior engineers; design team topology
- Represent technology to the board, investors and partners
An intelligent platform, built in layers.
The blueprint I adapt for AI-ready organisations. Raw data flows up through a semantic core, the ontology, so that every product, report and AI agent works from the same meaning. Platform and governance run the full height of the stack.
Applications change every year; what a "patient", "order" or "account" means should not. Put meaning at the core and you can swap models, vendors and frameworks without rewriting the business.
Agents grounded in a knowledge graph cite their sources, respect permissions and fail visibly. That's the difference between a chatbot and a system you can audit.
Security, compliance and operability bolted on later cost the most. Built in from day one, they become a delivery accelerator.
Scale is six problems, not one.
Throughput is only the start. I design for every dimension a growing organisation hits, and set explicit targets for each before a line of code is written.
Throughput
1M+ tx / dayEvent-driven, horizontally scalable services with idempotency and back-pressure built in.
Availability
multi-AZRedundancy, graceful degradation and tested recovery, with SLOs agreed with the business.
Latency
p95 budgetsLatency budgets per critical path, with caching and data placement designed to meet them.
Cost
unit economicsCost per transaction, per tenant and per AI token, measured and driven down as volume grows.
Security
zero-trustLeast privilege, encryption everywhere and auditable access, ready for regulated data from day one.
Organisation
team topologyPlatforms and boundaries that let more engineers ship independently without stepping on each other.
Prove it
- Architecture that won't need a rewrite
- MVP on managed services
- First AI use case in production
Make it reliable
- CI/CD, observability, on-call
- Domain model & data contracts
- First senior hires
Make it scale
- Event-driven re-architecture
- Internal platform on Kubernetes
- Knowledge graph & AI platform
Make it governable
- Multi-region, multi-tenant
- Compliance & audit at scale
- Architecture governance
One approach, many domains.
The method stays the same: model the domain, design for scale, build in governance. What changes is the vocabulary, the regulations and the failure modes, and I bring those with me so your team doesn't spend months learning them.
Healthcare & life sciences
Fragmented clinical feeds, competing code systems and strict privacy rules make every integration slow and every AI idea risky.
- Integrations that take months
- Research cohorts nobody can reproduce
- Clinical AI that legal won't sign off
- FHIR-first interoperability with terminology mapping
- OMOP CDM and OHDSI for research-grade analytics
- HIPAA-aligned platforms with de-identification and audit
Financial services & fintech
High volumes, zero tolerance for lost transactions, and auditors who want to see every decision.
- Batch jobs that can't keep up with growth
- Latency spikes at market open or month end
- AI pilots blocked by model-risk and audit
- Event-driven cores proven at 1M+ transactions a day
- Market and trading data pipelines
- Explainable, logged AI with model governance
Enterprise & B2B SaaS
Every team defines a customer differently, the platform buckles as tenants grow, and the AI assistant hallucinates.
- Data that never reconciles across teams
- Noisy-neighbour tenants and rising cloud bills
- Enterprise buyers asking for SSO, audit and residency
- Ontology and semantic layer as the shared model
- Kubernetes platforms with tenant isolation and FinOps
- Enterprise features: SSO/SCIM, RBAC, audit, regions
AI-native startups
Speed matters, but so does an architecture that survives the first big customer and the first due diligence.
- A demo that won't survive real users
- No senior technical leader yet
- Investors asking hard questions about the stack
- Founding architecture for agents, RAG and evals
- Fractional CTO cover until the right hire
- Investor-ready technical story and data room
E-commerce, retail & marketplaces
Search that can't find the exact SKU, order flows that break at peak, and personalisation built on messy product data.
- Search that misses exact product codes
- Peak-season order and inventory failures
- Product data inconsistent across channels
- Hybrid search combining vector and keyword retrieval
- Event-driven order and inventory pipelines
- Product ontologies and clean catalogue models
Media, content & knowledge platforms
Huge volumes of unstructured content that people, search engines and AI all need to find and trust.
- Content locked in silos and formats
- Taxonomies that drift over time
- AI answers with no sources
- Content knowledge graphs and taxonomies
- Aggregation and enrichment pipelines
- Grounded AI search with citations
Investors & private equity
You need to know whether the technology and the team can carry the plan, before you commit.
- Pitch decks that oversell the AI
- Hidden technical debt and key-person risk
- Unclear cost to scale
- Architecture, code, data and team assessment
- AI due diligence: what's real, what's wrapper
- Post-deal value-creation plans
Regulated & sovereign data (US & UAE)
Data residency, overlapping regimes and new AI rules shape what you can build and where it can run.
- Unclear rules across federal and free-zone regimes
- Data that must stay in-country
- AI use cases stuck in compliance review
- Region-pinned platforms with policy as code
- AI gateways with redaction, routing and audit
- Private LLMs for data that can't leave
The full stack, from meaning to machines.
Every layer below is one I've designed, built or operated in production. Breadth matters at this level: the costly mistakes happen at the seams between layers.
Strategy & leadershipWhere the technology meets the business
Ontology & semanticsA shared model of meaning across the enterprise
AI & agentsLLM systems that are measurable and governed
Machine learningModels, features and the infrastructure behind them
Data engineeringPipelines that carry the business, real-time and batch
DatabasesChoosing and tuning the right store for the job
Backend & systemsServices designed for throughput and change
Product & frontendThe surface people actually touch
Cloud & platformReproducible, secure, observable infrastructure
The right tool, chosen for the problem.
Polyglot by necessity. I choose languages for the job, the team that will maintain the system, and the ecosystem around it, not out of habit.
AI and LLM systems, data pipelines, ML, APIs, event processors, infrastructure automation.
Platform services, high-concurrency systems, Kubernetes tooling and distributable CLIs.
React and Next.js front ends, Node.js services, React Native mobile apps.
Analytical modelling, warehouse design, performance tuning across Postgres and BigQuery.
Enterprise services and healthcare research platforms (OHDSI WebAPI).
.NET services, algorithms and data structures.
Web applications, e-commerce APIs and content aggregation.
Angular and React applications, browser tooling, code generators.
Querying knowledge graphs and ontologies: property graphs and RDF.
Aggregation pipelines, schema design and query optimisation at scale.
Infrastructure as code across AWS, Azure and GCP.
Automation, CI/CD pipelines, Kubernetes manifests and Helm.
Case studies, described by the problem, not the logo.
Client work is confidential by default. What follows is representative of the problems I'm brought in to solve and the shape of the solutions.
One shared meaning for the whole business, ready for AI
A multi-division organisation where every system had its own idea of a customer, product and contract.
Reports never reconciled, and the first AI assistant confidently gave the wrong answers.
- Domain workshops with business owners to capture concepts, relationships and rules
- Formal ontology and controlled vocabularies, versioned and reviewed like code
- Entity resolution mapping CRM, ERP and documents into a knowledge graph
- A semantic layer serving BI, APIs and GraphRAG-based agents from the same model
- Metrics that reconcile across divisions
- AI answers that cite the graph and respect permissions
- New data sources onboarded by mapping, not rebuilding
- A model the business owns and evolves
From fragile batch jobs to an event-driven core at 1M+ transactions a day
A transactional platform built on synchronous calls and a shared database.
Peaks caused latency spikes, overnight batches overran, and releases were scary.
- Traced real bottlenecks with load tests and distributed tracing before changing anything
- Set SLOs and latency budgets for the critical paths with the business
- Moved hot paths to a partitioned event stream with idempotent, retry-safe consumers
- Redesigned MongoDB data access: aggregation pipelines, indexes and read models
- Sustained 1M+ transactions a day
- Predictable latency at peak
- Batch windows replaced by streaming
- Smaller, safer, more frequent releases
Audit-ready data and AI platform for regulated environments
A regulated organisation wanted to put its data and AI to work across the US and the Gulf.
Every new use case stalled in security reviews and data-residency questions.
- Mapped data flows, sensitive fields and obligations: HIPAA, SOC 2, GDPR and UAE PDPL
- Centralised identity and fine-grained access with policy as code (RBAC and ABAC)
- Built PII masking, encryption and region-pinned storage for residency
- Wrote immutable audit trails covering every human and AI-agent action
- Security reviews became a checklist, not a negotiation
- Clear data-residency story per region
- AI agents with auditable, permission-aware access
- Evidence ready for auditors on demand
From fragmented clinical feeds to research-ready, standards-based data
EHR, lab, claims and device data in different formats and code systems.
Each integration was a bespoke project, and research cohorts took months to assemble.
- Standardised exchange on HL7 FHIR, with adapters for legacy HL7 v2 feeds
- Normalised codes to SNOMED CT, LOINC and RxNorm through a terminology service
- Landed analytics in OMOP CDM with OHDSI ATLAS for reproducible cohort definitions
- Built in de-identification, consent and audit at every hop, then layered clinical AI on top
- Integrations measured in weeks, not months
- Cohorts defined once and reproducible
- HIPAA-aligned by design
- Work contributing to peer-reviewed research
Private LLMs for data that can never leave the building
Leadership wanted generative AI, but legal would not allow sensitive records to reach a public API.
The project had stalled at "we can't".
- Classified use cases by data sensitivity and the business value of each
- Stood up an AI gateway with authentication, PII redaction and cost controls
- Ran open-weight models on local GPU hardware for sensitive workloads, routing the rest to hosted models
- Grounded answers in the knowledge graph and gated every release on an evaluation suite
- Sensitive data never leaves your infrastructure
- Cost and quality tracked per use case
- A clear rule for what runs locally vs hosted
- A pilot legal could sign off
Client names, figures and specifics are withheld under confidentiality. Detailed references available on request.
What my background means for your project.
Experience is only useful if it changes outcomes. Here's how each part of mine turns into something concrete for you.
How I work with leadership teams.
Senior technical judgement is mostly about what not to build. These principles shape every engagement.
Model the domain first
A clear ontology of the business outlives any framework. Get the nouns and relationships right and the systems, analytics and AI fall into place.
Architecture is a business decision
Every technical choice is tied to cost, risk, speed or capability, and explained in those terms to the people who fund it.
Prove it in production
Thin vertical slices, shipped early, measured honestly. Demos persuade; production systems with evaluations and telemetry decide.
Own your stack
Avoid lock-in where it matters. Portable infrastructure, open standards and data you control keep your options open.
Leave the team stronger
Documentation, decision records and mentorship are deliverables. Success is the team running it confidently without me.
Discretion, always
Your roadmap, data and challenges stay yours. Confidentiality is the default.
Shaped around where you are.
From a focused two-week review to an ongoing seat at the leadership table. Every engagement starts with a no-obligation conversation.
Architecture & AI readiness review
An independent assessment of your systems, data and AI plans.
- Current-state audit
- Risk & opportunity map
- Prioritised roadmap
Fractional CTO
Executive technical leadership, part-time, for as long as you need it.
- Strategy & board reporting
- Hiring & team building
- Architecture ownership
Build & hand-off
Hands-on delivery of a critical platform, AI system or knowledge graph.
- Design through production
- Embedded with your team
- Documented transition
Advisory
A senior sounding board for founders, CTOs and investors.
- Regular office hours
- Technical due diligence
- Decision reviews
Research, community and open source.
Work that is public, peer-reviewed or open to inspection.
Science Immunology
Co-author on peer-reviewed research, applying engineering rigour to scientific data.
Read the paper →MongoDB Community Leader
Recognised contributor to the MongoDB community. Author of an open MongoDB cookbook and aggregation-pipeline tooling.
See the cookbook →80+ public repositories
AI pipelines, data tooling, infrastructure helpers, developer CLIs and full products, all open to review.
Browse GitHub →Writing on the problems I solve.
Practical notes on AI architecture, ontologies, data platforms, regulation and technical leadership. All articles →
Build vs buy for AI: a decision framework for leadership teams
Leadership teams keep asking the same question about AI: build it ourselves or buy it? The answer is almost never all of one. Here is how I break the decision down.
Explainer · Cloud & platform · AIGPUs on Kubernetes: what Dynamic Resource Allocation changes for AI platforms
For years Kubernetes treated a GPU as a number. Dynamic Resource Allocation treats it as a device with properties, partitions and health. For anyone running AI workloads, that changes how you share expensive hardware.
Explainer · Data platformsChange data capture: turning your database into a reliable event stream
Most companies already have their most valuable event stream. It is sitting inside the database's transaction log. Change data capture lets you use it without rewriting every application.
Analysis · UAE · AI infrastructureBuilding AI in the UAE: where should your models actually run?
The UAE is investing heavily in AI compute, and in-country cloud has matured. For teams building AI here, the hosting question is now less about availability and more about matching each workload to the right place.
What I'm building with right now.
Advice on AI is only as good as the hands-on time behind it. These days I'm running large language models on my own hardware, including an NVIDIA DGX Spark, to understand what local, private AI can really do for clients who can't send data to a public API.
Two minutes to a first diagnosis.
Answer four quick questions. You'll get the likely root cause, what the first 30 days would look like, and the engagement that fits. Then send it to me in one click if you want to talk it through.
Pick your answers on the left and your diagnosis will build here.
Common questions.
If yours isn't here, ask me directly. I'm happy to answer before any engagement.
What does a fractional CTO do?
A fractional CTO provides executive-level technology leadership part-time. I own technical strategy, architecture decisions, hiring and board-level communication, without the cost of a full-time hire, and hand over cleanly when you're ready for one.
Can you help us get our AI or LLM pilot into production?
Yes. I ground retrieval in a domain ontology and knowledge graph (GraphRAG), orchestrate agents with LangChain and LangGraph, and add evaluation, tracing and guardrails so quality and cost are measurable before you scale.
Do you work with healthcare data and HIPAA?
Yes. I design HIPAA-aligned architectures and work with HL7 FHIR, OMOP CDM and OHDSI, and clinical vocabularies like SNOMED CT, LOINC and RxNorm, with de-identification, consent and audit trails built in.
What is ontology engineering and why does it matter for AI?
An ontology is a formal model of the concepts in your business and how they relate. It gives every system, report and AI agent the same meaning, which makes data reconcile and makes AI answers accurate, explainable and permission-aware.
How do engagements start?
With a free 30-minute introductory call, which you can book directly from my calendar. Most clients then begin with a 2 to 4 week architecture and AI readiness review that produces a prioritised roadmap, then continue as fractional CTO, a build-and-hand-off project or advisory.
Do you work remotely with international teams?
Yes. I'm based in the United States and the UAE and work with teams worldwide, remotely and with on-site sessions when they add value.
From first note to kick-off, without the runaround.
No sales team, no gatekeepers. You talk to me directly from the first message, and you'll know within a week whether and how we should work together.
Book or write
Pick a slot on my calendar, or analyse your challenge first and bring the result.
Intro call
30 minutes, free. Your goals, constraints and whether I'm the right fit. Pick a time →
Proposal
A short written plan: scope, approach, timeline and cost. No 40-page decks.
Kick-off
Discovery starts right away, with a clear first milestone in the calendar.
Looking for an advisor? I'm up for it.
If you're building in AI, data, fintech, SaaS or healthcare and want an experienced technical voice on your side, as a board or startup advisor, a technical advisor to investors, or a sounding board for your CTO, let's talk.
Let's talk about what you're building.
The fastest route is a 30-minute call: pick any open slot. Prefer to write? Send a message and I'll reply personally within two business days.