Engineer

Software systems
and applied AI

My work covers AI-assisted development, security engineering, robotics and underwater acoustics. I co-founded Aimlink.ai and work on the architecture of its shooting sports apps.

Robotics
Robot arm test

Product engineering

Software for athlete training records, club operations and range scoring.

Engineering practice

AI &
project memory

I keep decisions, failed approaches and test results alongside the code. Later sessions can verify what was tried and resume from the evidence.

AI-assisted
development

I define the acceptance checks before an agent changes code. When an API changes, I test it through the app that calls it.

I split the work into investigation, disproof, implementation and review.

Inputs and review evidence
LocalProject notes

Decisions and past attempts

SharedEngineering notes

Decisions that apply across projects

Agent and tools

Inspect code, make changes, run tests

Review

Code changes, test results and open questions

Inputs and outputs of an AI-assisted code change.
The four rolesExplore → Disprove → Build → Judge
  1. 01ExploreEstablish the evidence

    Read the code and prior attempts. Record the question, evidence and acceptance checks.

  2. 02DisproveTest the explanation

    Look for counterexamples. Check whether the behaviour is intended or whether a fix already exists.

  3. 03BuildImplement the change

    Make a small, reversible change. Run the original case and the relevant regression checks.

  4. 04JudgeReview against the criteria

    Compare results with the acceptance checks. Keep, revise or revert the change and record why.

Research

Evaluating
AI-assisted work

I compare work with and without AI at the same quality bar. I look at time, review effort and what I can do independently afterwards.

  1. 01Long-term memorySource-linked project records

    I keep dated decisions in Markdown and Git. For a browser bug, I record the browser, viewport and observed failure so the next session can reproduce it.

    Read the MemGPT paper Read the LongMemEval paper
  2. 02Agent handoffsResuming interrupted work

    Record the task, scope and last verified state. After a tool failure, verify any partial changes before retrying.

  3. 03Learning with AIMeasuring independent performance

    Count time spent learning, correcting and reviewing. Use a later task without AI to test what was retained.

  4. 04Workflow evaluationTesting changes on fresh cases

    When an agent misses a source or misreads a result, I change one part of the workflow and compare both versions using the same checks.

    Read the Reflexion paper

Vulnerability assessment

Security
engineering

I review software and protocols for vulnerabilities. I develop a security scanner that uses AI agents to conduct assessments.

Vulnerability analysis
and verification

I reproduce findings, establish who can trigger them, assess impact and verify proposed fixes.

Review the design

Examine key handling and authentication guarantees.

Triage potential issues

Use scans and code review to identify cases worth testing.

Validate findings

Check reachability and safeguards. Record the evidence and remaining uncertainty.

Verify fixes

Retest the original case and check for regressions.

Discuss cybersecurity

AI security scanner

Assessment
workflow

AI agents explore the system, test suspected vulnerabilities and record the evidence behind each finding.

  • Agent step
  • Optional review
  • Reviewer step

Reviewer

  • Set scope

    Select the components and define the assessment goals.

  • Review evidence

    Inspect findings during or after the run.

  • Review for release

    Decide which findings and fixes are ready to release.

AI agents

  1. Confirm scope

    Read the assessment configuration.

  2. Explore the system

    Identify potential vulnerabilities.

  3. Test leads

    Choose and run checks to confirm or rule out each lead.

  4. Challenge results

    Have a second reviewer check the evidence.

  5. Compare findings

    Group related findings where the evidence supports it.

  6. Follow up

    Use the results to guide the next check. ↺ repeat

  7. Draft & retest

    Test candidate fixes in isolation when patch drafting is enabled.

  8. Report

    Record leads, confirmed issues and tested fixes.

Example assessment from configuration through release review.

Scanner
evaluation

Missed findings and false positives can become regression tests.

During a scan

Triage findings
  • Rank leads by evidence.
  • Merge duplicates.
  • Record why leads were dropped.

Remove sensitive details before saving test cases.

Use verified changes in the next scan.

Between scans

Test changes
  1. Explore Identify a missed case.
  2. Disprove Confirm the failure.
  3. Build Change one component.
  4. Judge Compare versions on fresh cases with the same budget.

Both versions use the same scoring criteria and run limits.

The evaluation holds the model fixed while testing changes to code, tests and instructions.

Scanner design
and status

Discuss AI security
  1. 01Run configuration

    The scanner applies the assessment configuration in code and logs each check.

  2. 02Reviewer controls

    Reviewers inspect the run log and sign off on results before release.

  3. 03Evaluation criteria

    I compare missed findings, false positives and cost across versions on cases held out of development.

  4. 04Assessment reports

    Reports document the cause, affected components and how to verify a fix.

  5. 05Implementation status

    The macOS scanner runs locally. The full workflow is in lab validation. Automated patching is at the design stage. Conventional scanners are separate tools in the broader security suite.

System diagrams

Engineering examplesMeasurement, robotics, AI and security

Example impact measurements

Impact measurements form one part of a result, alongside athlete, session and scoring data.

Eight example shots. A dotted line joins the two farthest shots, and a white cross marks their average position.
ES

Extreme spreadDistance between the two farthest shots

SD

Standard deviationSpread around the average position

Mean shotAverage impact position

The dashed ring illustrates spread in the sample.
Range data and scoring

The diagrams use examples and sample data.

About

Jens Bo Friis

My background is in cryptography, digital identity and software engineering. I’ve built desktop and mobile apps and APIs in Java, Python and JavaScript. At Applied Crypto, I work on systems that connect software to cameras, sensors and other hardware.

Technical background
AI agents

I develop tool use, shared project memory and handoffs between agent sessions.

AI-assisted development
Robotics

I work on robot arms, navigation and sensor feedback.

System diagrams
Cybersecurity

I review cryptographic code and protocols, validate findings and verify fixes.

Vulnerability analysis
Signal processing

I use filtering, spectral analysis and beamforming on multisensor data, including underwater acoustics.

Computer vision

I extract features and measurements from images, including target photos.

Range data and scoring
Software architecture

I design data models, APIs and application state for retries and partial failures.

Contact

Discuss
a project