Experiments
A collection of experiments.
Exploring AI capabilities, limitations, and practical usability. Each experiment is documented as a process: problem, hypothesis, experiment, failure, iteration, evidence, result, and lesson.
Status
Selected Experiments
All Experiments
The 35B MoE model runs on an older architecture with 12GB of VRAM.
Testing whether 12GB of VRAM and an older GPU architecture can run open-source LLMs optimally without putting undue strain on the GPU.
Read EXP 001: 35B MoE on 12GB VRAMPrefill speed on long context inference.
How can open-source LLMs achieve acceptable and reasonable prompt processing (prefill) speeds when the total context reaches 250K+ tokens?
Read EXP 002: Prefill Speed 256KAutomation agent workflow for business research needs.
What does it actually take to optimally use AI as an autonomous research assistant? Is it just the model? As it turns out, no.
Read EXP 003: Automation Agent WorkflowCreating digital employees with the help of open-source LLMs.
Is it possible to set up customer service, operational admin, ads specialist, data analyst, and finance functions using only workflows and open-source LLMs, given adequate hardware and reasonable resource usage?
Read EXP 004: Digital EmployeesBuilding a Full Feature Services Business Website Using Only Local AI.
Can a full-feature business website with before-after slider, page animations, 44-item cost calculator, and full SEO be built end-to-end using only local open-source LLMs without cloud AI?
Read EXP 005: Full Feature Website Local AIBuilding a Self-Contained Invoice System Without a Database.
Can an internal invoice tool with live A4 preview, customer autocomplete, and selectable text-based PDF be built with only local LLMs — without MySQL, just .js + 44 lines PHP on shared hosting?
Read EXP 006: Invoice System No DatabaseFrom 30 Minutes to 5 Seconds: Automating Mass CV Screening Without In-System AI.
Can 1–5 CV PDFs be screened from 30-minute manual review to 5-second ranked shortlist — without hallucination — using pdf.js, 24-node n8n, and 70% rule-based scoring on a private server?
Read EXP 007: Mass CV Screening 5sEach experiment is documented as a technical case study, showing not just results, but the process of thinking, testing, failing, and iterating.