Research. Benchmark. Build.
About HARBAN AI LABS
Mission
Make claims about AI checkable — by reading the research closely, recording how models are measured and by whom, and building on what the evidence supports.
Research.
Read the primary source. Explain what a paper did, what it showed — and what it did not.
Paper breakdownsBenchmark.
Record what each benchmark measures, where it breaks, and who produced every score.
BenchmarksBuild.
Turn that evidence into engineering practice: reference architectures, guides and, in time, services.
AI engineering
This site, by the numbers
Live counts of our own data
- Models tracked
- 14
- Benchmarks documented
- 14
- Benchmark results
- 35Each with its source
- AI tools listed
- 31
- Paper breakdowns
- 6
- Articles
- 4
- Videos
- 0None published yet
- Service lines
- 8All planned
What we research
Paper breakdowns are technical readings of influential AI papers, written in our own words: the problem, the method, what the paper reports, and where its evidence stops. Each one links the original paper and records the date we checked its details against it.
Research on this site is organised under a fixed set of topics:
- Agents
- Large language models
- Evaluation
- Retrieval (RAG)
- Fine-tuning
- Scaling
- Reasoning
- Architectures
- Prompting
- AI engineering
- Skills & careers in AI
- Tooling
- Safety & reliability
What we review and record
Benchmark scores are easy to quote and easy to misread. For each benchmark we record what it measures, its metric, its documented caveats and a dated health label — active, saturated, flawed or retired.
The model registry holds specifications confirmed on official pages, with a source for every filled-in field. Every result is labelled by who produced it and keeps the settings its source states, so you can see whether two numbers are comparable.
What we are building
Engineering is where the evidence gets used. Our AI engineering pages describe our engineering practice and reference architectures.
We are designing 8 service lines. All of them are planned — not yet offered — and each is labelled with its real status.
- LLM & model evaluationPlanned
- Reliability & red-team evaluationPlanned
- AI agent engineeringPlanned
- LLM application developmentPlanned
- RAG & knowledge systemsPlanned
- AI automation & data workflowsPlanned
- Prototypes & proofs of conceptPlanned
- Technical research & advisoryPlanned
There are no client case studies, logos or testimonials on this site. The published research, data and methodology are the evidence you can check today.
What we teach
For learners and practitioners: tutorials, analysis and engineering guides in Insights; explainers and walkthroughs in Videos (none published yet); a verified directory of AI tools; and a newsletter you can sign up for — no issue has been sent yet.
Principles
Provenance
Every external number links to the primary source it came from and the date we checked it. Every benchmark result says who produced it: the vendor, an independent evaluator, or us.
Independence of evaluations
Placement in the tools directory and position in any ranking on this site are not for sale. A result we measure ourselves must link to its run logs before it can be published.
Honest uncertainty
Unknown is shown as “Not verified”, never as a blank, a zero or a guess. Benchmarks carry health labels, so a score on a saturated or flawed test is read with that in mind.
Simplicity
Plain language, and the simplest method that answers the question — an exact check before a model-graded one, a small system before a platform.
Security
Collect as little as possible and protect what we keep. No analytics or advertising cookies; saved items stay in your browser; IP addresses used for abuse prevention are stored only as keyed hashes.
The full rules — sourcing, labels, corrections and our use of AI tools — are in our editorial and data standards.
Read the standards