About
A specialist Data and AI consultancy, led hands-on.
No pyramid, no bench, and no handover to a delivery team you have never met. The person who designs the system is the person who builds it.
The practice
Built on nearly two decades of doing the work.
Duspat is led by a practitioner with nearly twenty years across software engineering, cloud architecture, data platforms, integration and governed data delivery — spanning AWS solutions architecture, data and AI systems design, and large-scale migration and integration work.
That range is not a list of tools. It is the reason we can be specific about trade-offs. We have operated the platforms we recommend, paid the cost of the shortcuts, and had to explain an audit trail to somebody who genuinely needed it to hold up. Experience of that kind is not transferable from a slide deck.
The work has included AWS networking and hybrid connectivity for research environments, governed retrieval systems built on Amazon Bedrock and pgvector, event-driven platforms handling financial reporting at enterprise scale, and migrations off legacy warehouses with production tables that could not be taken offline. Alongside it: HLD and LLD architecture documentation, reusable Terraform modules, CI/CD standards and the reusable integration patterns that stop the fifth source system costing as much as the first.
Duspat exists because that combination — strategy, architecture, and the willingness to open an editor — is unusually hard to buy. Most organisations are offered one or the other, and then made to bridge the gap themselves.
- Nearly 20 years
In software, cloud and data
Software engineering, cloud architecture, data platforms, integration and governed data delivery.
- 8 sectors
Of domain exposure
From regulated research data to financial reporting, each with a different definition of "wrong".
- Advise → Govern
End-to-end capability
Strategy, architecture, hands-on build, deployment and the governance around it.
- Hands-on
Architecture and delivery
The person who designs the system writes the code. There is no pyramid and no handover.
Expertise
Six areas, and they compound.
The value is not in any one of them. It is that an architecture decision, a governance obligation and a deployment constraint can be reasoned about at the same time, by the same person, before any of them becomes expensive.
- Cloud architecture
- AWS solutions architecture across compute, storage, networking, eventing and integration — including hybrid and on-premises connectivity where the estate demands it. Azure and GCP where the estate points that way.
- Data platforms
- Lakehouse and warehouse design, dimensional and domain modelling, medallion architecture, and large-scale migration onto platforms a team can actually operate afterwards.
- Data & AI systems
- Retrieval-augmented generation, embeddings and knowledge-base patterns, agentic workflow design, and the evaluation and guardrails that let an AI system pass a security review.
- Integration engineering
- Event-driven architecture, API-led integration, streaming and batch pipelines, and reusable configuration-driven frameworks rather than the fifth bespoke job in a row.
- Governance & production readiness
- Access control, metadata, lineage, data quality, retention, auditability and monitoring — designed into the system rather than appended to it after an incident.
- Delivery engineering
- Infrastructure as code, CI/CD, reusable workflows, pre-commit checks, semantic versioning and the engineering standards that make the next release unremarkable.
Industry exposure
Experience includes work across eight sectors.
Each one has a different definition of “wrong” — a research environment, a royalty statement and a fuel reconciliation fail in entirely different ways. Knowing which kind of wrong you are guarding against is most of the architecture.
Sector-level only. No client is named, and none is implied.
Life sciences research
Sensitive research data, hybrid cloud and on-premises connectivity, and AI governance where the control model has to exist before anything ships.
Pharmaceutical-adjacent research
Regulated environments where auditability and access control are the design constraint, not a later compliance exercise.
Music publishing
Financial royalty reporting at enterprise scale, where every figure has to trace back to the event that produced it.
Aviation fuel management
Transaction data feeding reconciliation, billing and reporting — where an error is a financial error and must be traceable.
Retail
High-volume pipelines and integrations across systems that were never designed to talk to one another.
Travel
Event-driven integration and cloud automation across distributed, variable-load systems.
Financial reporting
Data platforms where accuracy is not negotiable and the audit trail is part of the product.
Digital platforms
Cloud-native services, APIs and data products built to scale with usage rather than with headcount.
Technical depth
The things we have actually used.
Published in full rather than curated, because if you are about to bet a roadmap on a platform, a tidy shortlist is not what you need from us.
- AWS
- Lambda
- API Gateway
- S3
- SQS
- SNS
- EventBridge
- Step Functions
- Glue
- Redshift
- Athena
- Kinesis
- DMS
- ECS
- DynamoDB
- IAM
- CloudWatch
- Transit Gateway
- Site-to-Site VPN
- Data platforms
- Databricks
- Snowflake
- BigQuery
- Apache Iceberg
- Delta Lake
- PostgreSQL
- Apache Hadoop
- MongoDB Atlas
- AI & retrieval
- Claude
- Amazon Bedrock
- RAG
- Embeddings
- pgvector
- Knowledge bases
- Pipelines & streaming
- Apache Airflow
- DBT
- PySpark
- Apache Kafka
- Spark Streaming
- Snowpipe
- Delta Live Tables
- AWS Glue
- Delivery & infrastructure
- Terraform
- CloudFormation
- Docker
- Kubernetes
- Helm
- GitHub Actions
- GitLab CI/CD
- AWS CodePipeline
- AWS SAM
- Serverless Framework
- Languages
- Python
- SQL
- TypeScript
- JavaScript
- Node.js
- Bash
- Java
- PHP
- Analytics & visualisation
- Power BI
- Apache Superset
- Looker Studio
- Kibana
Governance
Retrofitting governance is the most expensive way to buy it.
Classification, access, lineage, retention and audit are architectural decisions. Taken during design, they cost a conversation. Taken after go-live, they cost a re-platform.
The same holds for AI. A control model written before the model is chosen is the difference between a system that passes review and one that stalls in front of it indefinitely — which, in our experience, is where most AI pilots actually go to die. Not for want of a better model. For want of an answer to “what can it do, and what happens when it is wrong?”
Delivery values
Five commitments you could hold us to.
Each is written so that you could catch us failing it. That is the only test a value statement has to pass to be worth printing.
We write things down
HLD, LLD, decision logs and runbooks are deliverables, not afterthoughts. If the reasoning only exists in someone’s head, it does not survive their next role.
We design for the team that stays
If your team cannot operate it, we have not finished. The measure of an engagement is what works six months after we leave.
We are specific about trade-offs
Every architecture has a cost — in money, in complexity, or in the option it forecloses. We name it before you buy it.
We do the work
You speak to the person who designs and builds it. No pyramid, no bench, no handover to a delivery team you have never met.
We would rather tell you not to
If a simpler answer exists, or AI is the wrong tool for the problem, we say so — even when the larger engagement would have been ours.
Let’s talk about what you’re trying to build.
Bring a problem, an architecture, or an initiative that has stalled. The first call is technical, and you will speak to the person who would do the work.
30 minutes, no pitch deck.