Intelligence, applied to the work you already do.
Most useful AI work is not a model. It is retrieval, tooling, evaluation and the boring parts of integration done properly. That is the part we build.
JANNEX designs and builds applied AI systems: generative AI applications, retrieval over private knowledge, AI agents that use tools and trigger workflows, document intelligence, and automation across existing business systems. Engagements cover architecture, data preparation, evaluation and production operations.
The gap between a demo and a system
A prototype that answers well on ten questions is not the same as a system that answers acceptably on ten thousand, in front of customers, on documents nobody has cleaned since 2019.
- Answers that are fluent and wrong, with no way to tell which is which
- Knowledge scattered across drives, wikis, inboxes and a database nobody owns
- No evaluation set, so every change is a guess about whether quality moved
- Automation that works until the first exception, then quietly stops
- Cost and latency discovered in production rather than in design
How we build it
AI work fails on the edges, so we spend the effort there: what goes in, what is allowed out, and what happens when the model is unsure.
Ground it
Retrieval over your own content, with citations back to source, so an answer can be checked.
Bound it
Explicit tool permissions, allowed actions and escalation paths. The system asks a human when it should.
Measure it
A held-out evaluation set built from real questions, run on every change, before release.
Operate it
Logging, cost tracking, latency budgets and a route for correcting bad answers permanently.
What this covers
Generative AI applications
- Assistants embedded in your product
- Drafting and summarisation
- Classification and extraction
- Search that understands intent
- Multilingual handling
Knowledge and retrieval
- RAG architecture
- Chunking and embedding strategy
- Hybrid keyword and vector search
- Permission-aware retrieval
- Freshness and re-indexing
Document intelligence
- PDF and scan ingestion
- OCR pipelines
- Field extraction and validation
- Structured output with schemas
- Human review queues
Automation
- Workflow orchestration
- CRM and back-office actions
- API automation
- Exception handling
- Audit trails
Technology
The working set for this capability. Choices are made per engagement, against your constraints and your team's skills.
Where it is typically applied
Patterns we see repeatedly, described generically. Your version will differ in the details, and the details are the work.
Internal knowledge assistant
Staff ask questions in plain language against policy, product and process documents, and get answers with a link to the paragraph they came from.
Document processing
Invoices, contracts or forms read into structured fields, with anything below a confidence threshold routed to a person instead of guessed.
Support deflection
Common questions answered directly; anything else handed to an agent with the conversation and the customer record already attached.
Operations automation
A recurring multi-system process — quote, approve, update, notify — executed end to end with a record of every action taken.
How the engagement runs
The same seven stages, scoped to the size of the problem.
Understand
We start with the constraint, not the feature list. What breaks today, who it affects, what it costs.
Define
A written scope with the trade-offs made explicit — what is in, what is deferred, what we will measure.
Design
Interfaces, data models and system boundaries designed together, because they constrain each other.
Build
Short cycles against a working environment. Reviewed code, tests where they earn their keep.
Launch
Staged rollout with monitoring in place before traffic, not after the first incident.
Learn
Instrumented usage read against the thing we said we would measure at Define.
Scale
Performance, cost and operations tuned once real load has told us where the pressure is.
Questions we are asked
Do you train custom models?
Usually not, and usually you should not either. Retrieval over your own content plus careful prompting and evaluation covers the large majority of business cases at a fraction of the cost and with far easier updates. Where fine-tuning genuinely helps — narrow formats, tone, structured output — we will say so.
Where does our data go?
Wherever you decide. We design for the constraint: a specific region, a private network path, no retention by the provider, or a self-hosted model. That decision is made at architecture, not afterwards.
How do you stop it making things up?
Ground answers in retrieved source text, cite it, constrain output formats, and set a threshold below which the system says it does not know rather than filling the gap. Then measure how often each of those fires.
How long does a first system take?
A scoped, production-grade assistant or automation typically runs weeks rather than months. The variable is almost never the model — it is the state of the data and the number of systems it has to touch.
Related reading
AI chatbots vs AI agents: what businesses actually need
The two are engineered differently, fail differently and cost differently. Choosing the wrong one is the most common way an AI project stalls.
Read AI · 6 min readRAG vs fine-tuning: which approach makes sense?
They solve different problems. Most business cases need retrieval; a minority genuinely benefit from fine-tuning, and a few need both.
Read Automation · 6 min readWhich business processes should you automate first?
A ranking method that survives contact with reality, and the three characteristics that make a process a bad first candidate.
ReadHave something worth building?
Tell us the constraint you are working against. If we are not the right people for it, we will say so.
Or write to connect@jannex.in