Retrieval over your own documents
Answers grounded in your contracts, tickets and manuals — each one citing the paragraph it came from, so a wrong answer can be traced instead of argued about.
pgvectorHybrid searchClaude API
Applied AI in production · 14 countries since 2019
Enterprise software, engineered to last.
We design, build and run business-critical systems — core platforms, cloud modernisation, data infrastructure and the applied AI inside all three. Forty-two engineers and designers, one delivery structure.
- Fixed-fee discovery
- You own all the code
- SLA-backed support
p95 42 ms · 0 incidents in 90 days
A support-triage agent running in a client’s production estate — it reads the ticket, calls the systems of record, and drafts the resolution for a human to approve.
- 0
- Systems delivered
- 0
- Engineers & designers
- 0
- Client retention
- 0
- Countries served
Built and operated on
- AWS
- Google Cloud
- Azure
- Cloudflare
- Kubernetes
- PostgreSQL
- Snowflake
- Claude
- Redis
- Docker
Applied AI
AI that has to justify itself
We have built AI into systems that process claims, price freight and answer regulated customer questions. The engineering is the easy half. The hard half is proving the thing is right, keeping it right as the data moves, and knowing which problems should never have had a model pointed at them.
What one request goes through, in every AI system we ship
- Question from your app or a person
- Retrieval your documents, with permissions applied
- Model chosen per task, with a cost ceiling
- Tools your systems of record, behind an approval gate
- Verified answer cited, scored, and logged
Agents that do the work
Multi-step agents that read a case, call your systems and finish the task — with a human approval gate on anything that moves money, data or a customer’s status.
LangGraphTool useCelery
Forecasting and prediction
Demand, churn, credit risk and capacity models trained on your own history, retrained on a schedule, and always reported against the baseline you use today.
PyTorchdbtAirflow
Document and image intelligence
Invoices, claims, KYC packets and scanned forms turned into structured fields — with a confidence score, and a review queue for the ones a person should still see.
OpenCVOCRTriton
Support and internal assistants
Assistants over your product, policies and knowledge base that answer in your own wording — and hand over to a person the moment they are out of their depth.
StreamingNLPClaude API
Evaluation, guardrails and cost control
Every model we ship carries an evaluation suite, a drift alarm, a prompt-injection boundary and a monthly cost ceiling. Without those it is a demo, not a system.
W&BvLLMGrafana
Chosen per task, not per contract
A frontier model for reasoning, a small fine-tuned one for the high-volume path, a classical model where a classical model wins. Most systems we build use more than one, behind an interface that lets you swap them.
ClaudeGPTLlamaMistralEmbeddingsFine-tuned SLMs
In your account, under your policy
Managed endpoints on Bedrock, Vertex AI or Azure AI Foundry, or self-hosted on your own GPUs with vLLM when residency or cost demands it. Your data stays in your account, and is not used to train anyone’s model.
BedrockVertex AIAzure AIvLLMOn-prem GPU
Not every problem needs a model
If a rule, a query or a better form solves it, we will tell you that and build the cheaper thing. A model you cannot evaluate is a liability you pay for monthly, and we would rather lose the line item than hand you one.
What we do
Nine practices, one delivery team
Engineering, design, quality, infrastructure and support report into the same programme, on the same plan, accountable for the same release date.
Product Engineering
New products and platforms your own team can maintain after handover — typed end to end, tested, documented.
TypeScriptNext.jsNode
Applied AI
Retrieval, agents and models placed where they measurably reduce work — each with an evaluation suite and a cost ceiling.
Claude APIRAGPyTorch
Cloud & DevOps
Infrastructure as code, zero-downtime releases, autoscaling, and a cloud bill that stops outgrowing revenue.
AWSKubernetesTerraform
Legacy Modernisation
Incremental replacement of systems you cannot switch off — dual-run verification, no big-bang cutover.
.NETJavaPHP
Mobile Applications
iOS and Android from one codebase, with offline-first sync, on-device inference and store release automation.
FlutterSwiftKotlin
Data Engineering
Pipelines, warehouses and a semantic layer everyone trusts, so every metric has exactly one definition.
dbtAirflowBigQuery
Quality & Test Automation
Regression suites, load and security testing, and CI gates that stop a bad release before your users find it.
Playwrightk6OWASP
Managed Support
Monitoring, patching and an on-call rota against a written SLA — for what we built and what you inherited.
24×7SLAGrafana
Dedicated Teams
Named senior engineers embedded in your organisation, reporting into your leads, on a monthly rolling term.
No benchYour process
Why JupiterMint
Commitments we put in the contract
Every firm you shortlist will claim quality and good communication. These are the checkable undertakings we sign up to instead.
You own everything
Code, cloud accounts, model weights and documentation sit in your organisation from the first commit.
No subcontracting
The people named in the proposal write the code. We do not resell someone else’s bench.
Fixed-fee first phase
Discovery is a defined scope at a defined price, ending in a plan — not an open-ended meter.
Fortnightly demos
A working build every two weeks. Status is an environment you can open, not a slide we send.
No unmeasured AI
Every model in production has an evaluation suite, a cost ceiling and a defined failure behaviour.
Support against an SLA
Response and resolution targets written down, with the credit that applies when we miss them.
Our work
Products we built, and still run
Own product · SaaS
Live at cintram.com
Cintram — client delivery platform
For service businesses that deliver structured client work over weeks or months: a closed deal becomes a project with its phases already in place, and the client follows progress in a white-labelled portal instead of chasing status emails. Around twenty product modules — CRM, sales pipeline, projects and planner, LMS, marketing, forms and billing — behind one OpenAPI-typed contract shared by three surfaces.
- 20+
- Product modules
- 3
- Client surfaces
- 15 → 22
- Angular
- 4.2 → 5.2
- Django
Angular 22Django 5.2DRFCelery
PostgreSQLRedisOpenAPIDocker
Financial services
Credit underwriting copilot
Retrieval over twelve years of filings and internal policy, drafting the credit memo with a citation on every assertion.
- -71%
- Review time
- 11 wks
- To production
Remote patient monitoring
Forty clinics off phone calls and spreadsheets, with on-device triage that works without a signal.
- 1.2M
- Patients
- -46%
- Readmissions
Fleet routing platform
Twelve thousand vehicles re-planned every thirty seconds against live traffic and fuel prices.
- -38%
- Fuel cost
- -61%
- Compute
Technology
The stack we work in
Chosen for the problem, not the résumé. If your team already runs .NET on Azure we will make that good rather than sell you a rewrite.
- React
- Next.js
- TypeScript
- Angular
- Vue
- Tailwind
- Vite
- Storybook
- Node.js
- Python
- Go
- Laravel
- .NET
- Java
- GraphQL
- Celery
- Flutter
- Swift
- Kotlin
- React Native
- TF Lite
- Fastlane
- PostgreSQL
- ClickHouse
- Snowflake
- BigQuery
- Kafka
- dbt
- Airflow
- Redis
- AWS
- Google Cloud
- Azure
- Kubernetes
- Terraform
- Docker
- Cloudflare
- Grafana
- Claude API
- PyTorch
- LangGraph
- pgvector
- vLLM
- Triton
- W&B
- OpenCV
Domains
Where the domain knowledge already exists
Ten domains we have shipped into more than once — so the rules, the integrations and the awkward edges are already known rather than discovered on your budget.
Financial services
Healthcare
Logistics
Education
Manufacturing & IoT
Public sector
CRM
POS & billing
Tracking systems
Gaming
Client references
They shipped, then they stayed
“We had been through two vendors before JupiterMint. They replatformed us in eleven weeks and it has not had an unplanned outage since.”
“The triage agent paid for the whole engagement inside the first quarter. What convinced our board was the evaluation report, not the demo.”
“Fortnightly demos, honest estimates, no surprises on the invoice. They behave like an internal team that happens to be very good.”
FAQ
Before you get in touch
Anything else, just ask — you will get a reply from an engineer within one working day.
You do, from the first commit. Repositories, cloud accounts, model weights and documentation sit in your organisation throughout. There is no proprietary framework you have to keep paying us for.
Assessment work typically begins within five to seven working days of signature. Where it is genuinely urgent we can place two senior engineers within forty-eight hours while the full team is assembled.
Most of our engagements are structured that way. We join your board, your standups and your review process, and knowledge transfer is written into the plan.
Every AI feature ships with a token budget, response caching and model routing — a smaller model handles the request first and escalates only when needed. You set the per-request ceiling.
Thirty days of hypercare is included with every build. Most clients then move onto a managed support agreement covering monitoring, patching, an on-call rota and a monthly improvement allowance.