AI security & evaluationExperimental open source
Prompt Injection Detector
Exploring a lightweight classification layer for untrusted text in LLM, retrieval, and agent workflows.
Explore case studyINDEPENDENT AI ARCHITECTURE & DEVELOPMENT
Intelligent systems.
Real-world purpose.
Agentic AI systems built around your business workflows. From the first architecture decision to a working, evaluated pilot.
Ideas are good.
Inspectable work is better.
Inspectable work is better.
workflow.blueprintv0.1
ILLUSTRATIVE ARCHITECTUREHUMAN IN THE LOOP
01 / UNDERSTANDA question. A task. A goal.
context in, clarity out
Retrieve knowledgeGrounded in your sources
Use the right toolsBounded, purposeful actions
03 / EVALUATECheck. Validate. Ask a human.
A useful, reviewable outcome
DESIGNED FOR OVERSIGHTLess magic. More method.
FOCUSED ON WHAT MATTERS Agentic workflows Retrieval systems Evaluation & oversight Practical integration
THE STARTING POINT
A business problem. A practical solution.
AI is useful when it makes a real workflow work better.
01
Knowledge that’s hard to find
Connect scattered documents to answers people can trace back to a source.
02
Work that keeps repeating
Connect the right tools to a bounded workflow, with a person in control.
03
A prototype you need to trust
Find failure modes, define useful tests, and make the next decision with evidence.
HOW I CAN HELP
From the right questions to the right build.
AI architecture & feasibility
A clear path from idea to implementation.
- Workflow map and architecture options
- Risk register and evaluation plan
Agentic AI workflows
Purposeful agents. Bounded actions. Human control.
- A bounded workflow pilot
- Tool integrations and human review points
RAG & knowledge systems
Make your knowledge useful, with sources attached.
- Ingestion and retrieval pipeline
- Citations and an evaluation dataset
SELECTED WORK
Less promise. More proof.
Open, inspectable projects—with the context and limitations included.
Applied machine learningResearch / educational
Heartbeat Anomaly Detector
A research and educational audio-classification project connecting signal processing, a neural model, and integration guidance.
Explore case studyA CLEAR WAY FORWARD
Small steps. Shared understanding.
01
Discover
Understand the problem, the people, and the constraints.
02
Scope
Choose a useful outcome and agree how to measure it.
03
Build & evaluate
Implement, test, inspect failures, and iterate.
04
Hand over
Leave clear documentation and a practical path forward.
THE THINKING BEHIND THE BUILD
Technical curiosity.
Practical engineering.
Kalscript is an independent AI consulting practice built around clear thinking, inspectable work, and systems that keep people in control.
More about the approachFIELD NOTES
Ideas worth putting to work.
01
Retrieval & evaluation / 3 MIN READ
How to evaluate a document Q&A assistant
A useful answer needs more than fluent wording. Check retrieval, evidence, and the system's ability to say it does not know.
Read insight02
Architecture / 3 MIN READ
When does a workflow need an AI agent?
Start with the decisions a workflow needs to make. Then choose the smallest useful amount of autonomy.
Read insightLET’S BUILD SOMETHING USEFUL
Start with a problem.
Let’s find the right approach.
Bring your idea, your existing prototype, or a workflow that could work better.