HLab
Results, not adjectives.
Interactive views of numbers published in the project repositories. Every chart names its source, and every dataset is available as a table.
L.1 · TuneLab
Did fine-tuning actually generalize?
Every LoRA run is scored against the frozen base model on the same prompts. The gap between the test split and the held-out phrasings shows how much was memorized versus learned.
Data provenance. Synthetic ticket-triage task, n = 120 per split, one seed per configuration. Base model: Qwen2.5-0.5B-Instruct (4-bit). Source: tunelab-llm-finetuning · README.md, docs/RESULTS.md ↗
View data table
| Run | Test exact match | Held-out exact match | Held-out priority | Train minutes | Trainable params (M) |
|---|---|---|---|---|---|
| Base model | 8.3% | 5.0% | 46.7% | - | - |
| LoRA r8 · 300 steps | 100.0% | 63.3% | 69.2% | 9.7 | 2.93 |
| LoRA r8 · 100 steps | 100.0% | 50.0% | 50.0% | 4.4 | 2.93 |
| LoRA r16 · dropout 0.1 | 99.2% | 40.0% | 40.0% | 10.1 | 5.87 |
L.2 · AEGIS AI
Catching fraud without flooding analysts
A good fraud model alerts on almost every attack pattern while staying quiet on legitimate edge cases like travel or new devices. Compare alert rates by scenario, the two candidate models and the final confusion matrix.
Data provenance. Synthetic transactions. Test split: 25,877 transactions, 564 fraudulent (2.18%). Threshold 0.432. Source: aegis-ai-real-time-fraud-platform · README.md, docs/ML_MODEL.md ↗
View data table
| Scenario | Type | Alert rate |
|---|---|---|
| Impossible travel | Fraud | 100% |
| Card testing | Fraud | 98.7% |
| Account takeover | Fraud | 83.6% |
| Collusive merchant | Fraud | 81.8% |
| High-value anomaly | Fraud | 60% |
| Legit new device | Legitimate | 8.8% |
| Legit large purchase | Legitimate | 0.95% |
| Legit travel | Legitimate | 0.57% |
| Normal | Legitimate | 0.3% |
| Logistic regression | Model | P 0.564 · R 0.601 · F1 0.582 · PR-AUC 0.614 · ROC-AUC 0.976 |
| Hist. gradient boosting | Model | P 0.842 · R 0.885 · F1 0.863 · PR-AUC 0.926 · ROC-AUC 0.997 |
| Confusion matrix | HGB @ 0.432 | TN 25219 · FP 94 · FN 65 · TP 499 |
L.3 · WorkMesh
Where does permission-aware latency go?
Every WorkMesh read applies authorization in SQL under row-level security. This benchmark shows the spread from median to tail latency per endpoint. The dashboards that aggregate across teams have the longest tails.
Data provenance. 300 requests per endpoint at concurrency 10 on seeded data (51 people, 12 projects, 98 work items). Local machine, zero errors. Source: workmesh-company-visibility-platform · README.md ↗
View data table
| Endpoint | p50 (ms) | p95 (ms) | p99 (ms) | req/s |
|---|---|---|---|---|
| Health | 15.8 | 59.5 | 108.9 | 478 |
| Employee dashboard | 80.2 | 101.5 | 174.2 | 118 |
| Team Pulse | 68.9 | 121.7 | 138.3 | 129 |
| Leadership dashboard | 83.3 | 164.2 | 252.3 | 105 |
| Project list | 32.7 | 38 | 40.5 | 298 |
| People graph | 54 | 63.7 | 68.5 | 182 |
| Work graph | 39.8 | 53.7 | 57.6 | 242 |
| Update feed | 31 | 37.5 | 39.4 | 314 |
| Search | 54.9 | 62.6 | 88.2 | 177 |
L.4 · SourceAtlas
Does verified RAG hold up with a real model?
SourceAtlas is evaluated twice: with deterministic fixtures and with a live local model. Live answer correctness climbed across three evaluation runs as the pipeline was fixed.
Data provenance. 16-question synthetic evaluation set. “Live” uses qwen3:4b + nomic-embed-text; “Fixture” uses deterministic stand-ins. Source: sourceatlas-agentic-rag · README.md, eval/results/ ↗
View data table
| Metric | Live | Fixture |
|---|---|---|
| Answer correct | 92.3% | 84.6% |
| Citation correct | 95.8% | 69.2% |
| Abstention | 87.5% | 93.8% |
| Conflict detection | 100.0% | 100.0% |
| Injection handling | 100.0% | 100.0% |
| Live run 1: answer correct | 53.8% | - |
| Live run 2: answer correct | 69.2% | - |
| Live run 3: answer correct | 92.3% | - |
Tableau Public
Dashboards on Tableau Public.
Published on my Tableau Public profile. They open on Tableau's site.