Documents to reports
Turn source material into a structured draft with a clear output format, ready for someone to review.
Start with the requirementsAn independent agent software studio
AI agents,
built for work.
For teams bringing agents into their existing workflows.
Clear tasks. The right tools. Changes you can check.
Starting with MerlinThe agent skill layer
For developer teams.
Intended workflows.
Make repeated work
easier to hand off.
Reading documents, preparing reports, checking a repository. Bringing agents into that work means deciding what they can use, what they should produce, and when a person needs to review.
We’re starting with the skill layer: the reusable instructions and tools behind each task. These are the workflows we want to support as that foundation develops.
Task to tool
Match a task to a document reader, report writer, or repository check with the access and outputs it needs.
Explore skill selectionChange to review
Keep proposed skill changes separate, check example outputs, and retain the reason behind a revision.
Explore change checksOur first product · Research preview
Merlin.
The skill layer, managed.
A prototype for choosing skills, checking proposed changes, and keeping a versioned record. It works around the agent’s existing model.
01 / SELECT
The right skill.
For the right task.
As a team’s skill library grows, choosing by name gets harder. Merlin compares task requirements with what each skill declares, so missing information stays visible before selection.
The research behind MerlinChoose a taskINTERFACE CONCEPT · SAMPLE DATA
Create report.json with a summary.
- Report writerStructured reports with a summarySelected
- PDF readerRead documents and reply in chatNo fit
- Test runnerRepository checks with shell accessNo fit
- General helperNo explicit contract declaredUnknown
A declared match is a reason to select a skill. It isn’t proof of performance.
02 / CHECK
Improve a skill.
Keep the working
version intact.
A proposed fix becomes a separate candidate. Check its output, keep the active version intact, and give the reviewer a record of what changed and why.
Try editing either sample. The record follows your changes.
Where we’re taking itCheck a proposed changeLOCAL SAMPLE CHECK
v1 · active, unchangedv2 · candidate
✓ Sample passed
The summary field is present and non-empty.
× Sample failed
The summary field is missing.
This page only checks a non-empty summary field in your browser. It does not run a skill or provide a security sandbox.
Official result · Not evaluatedPromotion · Not evaluated
Version record
v1 · Active version. Kept intact.
v2 · Candidate based on v1. Reason: require a summary.
Sample evidence: A passed · B failed.
Official result and promotion: Not evaluated.
Planned focus
Built for the way
a team works.
The next direction: custom agent workflows for developer teams, backed by reusable tools to select, test, and maintain their skills.
Build around the work.
Start with one recurring task, connect the tools a team already uses, and define the output and human review points.
Keep it working.
Develop Merlin into reusable software for maintaining the skill library: requirements, candidate changes, checks, and a record a team can inspect.
Starting point: Merlin’s research prototype. Next step: validate these workflows with real team requirements.
Research behind the software
Practical software.
Research underneath.
Won Jeong is the founder of Londinium, an independent AI agent software studio. His work focuses on the systems around an agent: selecting the right tools, checking proposed changes, and keeping an inspectable record. Through Merlin and research in Korean legal question answering, efficient language models, and edge vision, he explores how research ideas can become practical software for real workflows.
Selected research
KDA Transfer
In three small language models, lower teacher-forcing loss after post-hoc KDA substitution did not restore autoregressive generation.
EdgePPE-Net
A prototype PPE-checking pipeline on Raspberry Pi 5 and Coral Edge TPU using conditional pose estimation.
LegalHarness
Research into Korean legal question answering, including retrieval coverage and answer control. Manuscript in preparation.
Separate research projects, not components of Merlin.
Start with a task.
Build the right agent.
Explore the skill layer we’re building
for more manageable agent workflows.