I build small, honest, measured LLM-efficiency tools. Each project below is a fresh implementation with a novel delta, tested on CPU, with receipts + charts. Prior art is credited in every README — no silent forks.
→ https://dastanzar.github.io/portfolio-novel/ (from portfolio-novel/docs/)
| Motion | What |
|---|---|
 |
bandit-spec tracking oracle K |
 |
block-evict freeing whole blocks |
 |
gate-keep memory beats drop |
Videos: record 30s per build with ScreenToGif / OBS → docs/assets/*.mp4 → embed as <video> on Pages + link here. YouTube: upload playlist “LLM Efficiency ×10” and add thumbnails below.
Images link to full PNGs. Numbers from
receipts/receipt.json(reproduce:pytest + demos/demo.py).
| # | Build | Result | Working | Implementation | Applications |
|---|---|---|---|---|---|
| 1 | kurt-switch  |
79.9% MSE cut | kurtosis-switched 3-codebook + LS scale | compute_row_stats, select_codebook |
Q4 on-device serving |
| 2 | shell-vq  |
64.5% vs flat | radius shells + per-shell scales + 1-bit residual | quantize_shelled, residual_refine |
sub-2-bit compression |
| 3 | dual-bias-fp8  |
2× accuracy @ same BW | per-block E5M2/E4M3 + bias 11–19 | choose_format, gemv_reference |
no-FP8 GPUs |
| 4 | bandit-spec  |
regret 13.7 | Thompson over K + p99 SLO veto | thompson_pick, slo_guard |
SLO chat serving |
| 5 | fair-batch  |
Jain 0.950, p99 −13% | DRR + DVFS + KV admission | simulate_fair_continuous |
multi-tenant |
| 6 | block-evict  |
frees whole blocks | whole-block scoring, COW intact | BlockCache.block_evict |
vLLM long-ctx |
| 7 | gate-keep  |
+11.8% @ b8 | ridge-gate + 8-d latent memory | fit_gate_ridge, compress_with_memory |
32k chat |
| 8 | epi-norm  |
52.8% HBM saved | Norm+residual+SwiGLU fused, recompute | fused_epilogue_ref, hbm_accountant |
training kernels |
| 9 | cache-roof  |
81.5% of peak | L2-split + dequant-overlap + CIs | cache_adjusted_roof |
kernel pick |
| 10 | green-roof  |
$3.59/1M, 393g CO₂ | $/J/CO₂ + autosolver + fair_split | replicas_for_slo |
capacity quotes |
📦 How each is structured (portfolio standard)
<build>/ {core package, tests/ (5–8 passing), demos/demo.py → receipts/receipt.json + assets/*.png, README with attribution, pyproject, LICENSE}
git init portfolio-novel && push→ create 10 repos (or one monorepo) + enable Pages from/docs- Create repo
DastanZar/DastanZar→ paste this file asREADME.md - Replace
dastanzar.github.ioURLs after first Pages deploy; add MP4s + YouTube links - Archive the 10 old templated repos (Settings → Archive) so portfolio shows only novel work