Published casesIncluded
Public experienceIncluded
Private files and recruiting dataExcluded
The source boundary comes first. The agent can navigate published work, but it cannot search private files or invent claims.
Portfolio project · agentic research
Approved evidence set
Portfolio Research Agent
I designed the public source boundary, research path, and evaluation standard for this agent. It turns a visitor’s question into a small research task, shows the evidence behind its answer, and refuses claims the published portfolio cannot support.
What the agent provides
A researched path to the right project
It interprets the question, makes a small search plan, compares approved passages, and directs the visitor to the best-supported source.
Approved public sources onlyNo private files, raw logs, or attachments
Visible uncertainty pathIt asks, narrows, or refuses when evidence is missing
Prototype boundary: this case documents an agentic research interaction and evaluation contract. It does not connect to recruiting systems, rank candidates, retain visitor questions, or claim model-quality results before the source set and evaluation run are complete.
01 / Context
A useful agent needs a clear boundary before it takes its first research step
This agent is designed to navigate a public portfolio, not to act as a general-purpose chat box. Its source boundary freezes only publishable material, then keeps each answer linked to the page and section that supports it.
Included
Approved About content, public experience facts, and published case studies.
View approved portfolio context →Excluded
Private documents, raw source files, personal attachments, backup files, and any recruiting data.
Answer boundary
The assistant narrows or refuses a question when the available evidence is missing, ambiguous, or only partially supports the claim.
02 / Solution
The agent plans a small search, checks the evidence, then guides the visitor forward
- 01Interpret
Understand the question
Identify whether the visitor needs a case, a specific fact, or a comparison.
- 02Plan
Choose a search path
Decide which approved projects and passages are most likely to help.
- 03Research
Read a small evidence set
Retrieve only relevant, traceable passages instead of the entire portfolio.
- 04Verify
Check every claim
Map each factual statement to the selected evidence before replying.
- 05Guide
Recommend the next page
Answer with sources, ask for clarification, or refuse what the public evidence cannot support.
03 / Evidence
A visible research trace makes an agent’s answer easier to trust
This browser-only demonstration renders prepared outcomes for supported, ambiguous, and unsupported example questions. The trace illustrates the proposed agent behavior. Your question stays in this browser and is not saved.
Try the local prototype
Prepared agent trace
How does Mônica keep AI workflow output reviewable?
PlanCompare review and handoff patterns
EvidenceCatalog, delivery, and support cases
Next stepOpen the supporting projects
The published cases keep human review in the workflow: catalog drafts are sent for review, delivery exceptions are handed to a person when uncertain or high-impact, and consequential support cases stay with a person. This prototype cites those published pages rather than inferring a broader claim.
04 / Evaluation design
I defined how this agent should earn trust before it is put into use
Freeze the test cases
I prepared answerable, multi-source, ambiguous, unsupported, and adversarial questions, with the expected evidence or refusal path for each.
Measure the right behavior
The evaluation contract separates source coverage, groundedness, correct refusal, latency, and failure type. Results remain unclaimed until the frozen set is run.
Keep the record minimal
The proposed telemetry stores only what supports quality review: question category, source IDs, validation result, and latency, not unnecessary visitor text.