# Angel Marinov / 0xamlab — Technical portfolio

Portfolio: https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/
GitHub: https://github.com/0xamlab
LinkedIn: https://www.linkedin.com/in/aimarinov/

Evidence snapshot: 2026-11-09. Page updated: 2026-11-10.

## Technical contributions

I build and optimize the parts of AI systems where execution cost accumulates: inference paths, GPU kernels, search, and runtime internals.
Selected work includes persistent vLLM KV-cache reuse, compiled Mojo retrieval and scientific kernels, and proposed upstream changes to quantized matrix multiplication and phrase-query memory.

I use coding agents for investigation, implementation, and test iteration. Architecture, engineering decisions, validation, and final review remain my responsibility.

## Evidence boundaries

All results are author-published and workload-specific. Benchmarks have not been independently rerun. Open pull requests are proposed contributions; statuses here reflect the evidence snapshot, not a current upstream-status check. The interactive animation is a schematic, not telemetry. No overall score combines latency, throughput and memory.

## 1. Persistent KV reuse

Category: Inference
Result: 4.38× batch wall-time speedup.
Workload: A40 · Qwen2.5-7B · BF16 · 16K prefix · concurrency 8
Comparison: versus forced prefix recomputation
Evidence status at snapshot: Published benchmark
Contribution: Built the external cache connector and its benchmark / lifecycle validation.
Detail: Fresh-process persistent-cache hits reduced mean batch wall time from 23.655112 s to 5.396244 s. Pooled TTFT p50 changed from 12.908762 s to 2.706828 s. The separate native-hot arm measured 2.418234 s mean wall time.
Limitations: Same-process native-hot cache was faster: 2.42 s. This is a persistent-cache result, not a blanket vLLM speedup. Three repetitions per cell; no statistical-significance claim.
Validation scope: Published validation includes 25,414 / 25,414 16K blocks and 2,949 / 2,949 4K blocks. Bitwise restored-KV and 64-token output equality are scoped to batch-invariant mode and greedy decoding.
Evidence quality: Source reviewed; benchmark not independently rerun.
Portfolio section: https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-cache
Recorded date in source: 2026-11-08
Benchmark record: https://github.com/thyn-ai/telys-vllm/blob/main/README.md
Detailed results: https://github.com/thyn-ai/telys-vllm/blob/main/bench/CONCURRENT.md

## 2. Quantized matrix multiplication

Category: Kernels
Result: 38.1% lower latency in the best tested cell.
Workload: M4 Max · 4-bit weights · BF16 activations · M=2048
Comparison: versus the pinned upstream main branch
Evidence status at snapshot: Open upstream PR
Contribution: Authored the proposed tile-dispatch optimization and regression coverage.
Detail: N=12288, K=4096, group size 64. The patch adds 64×64 tiles for selected non-NAX, non-batched affine QMM paths while leaving other dispatch paths unchanged.
Limitations: Five tested configurations report 5.6–38.1% lower latency. Kernel-level measurements, not end-to-end model speedups. Submitted; not merged.
Validation scope: The PR reports 982 tests passed / 58 skipped, plus 90 quantized-matmul subtests.
Evidence quality: Author-reported PR measurements; upstream acceptance not established.
Portfolio section: https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-gemm
Recorded date in source: 2026-11-03
Patch and measurements: https://github.com/ml-explore/mlx/pull/4826

## 3. Fuzzy search at 100K documents

Category: Retrieval
Result: 38.7× warm-query speedup in the selected cell.
Workload: M4 Max · 100K docs · 16-code-unit patterns · warm
Comparison: versus Fuse.js 7.1.0
Evidence status at snapshot: Published benchmark
Contribution: Built the native acceleration path and API-compatible integration.
Detail: Fixed-seed corpora with 15 patterns per cell. The source reports identical result order and zero score difference in the tested corpus. Cold index-build-plus-first-query measurements are separate.
Limitations: Warm results range from 8.8× to 38.7× across nine cells. The accelerated configuration uses 16 threads; this is not a single-thread comparison.
Validation scope: The published corpus reports exactly zero score difference and identical result ordering. This does not assert support for every library option.
Evidence quality: Author-published synthetic-workload benchmark; not a universal search-speed claim.
Portfolio section: https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-fuzzy
Recorded date in source: 2026-09-19
Benchmark record: https://github.com/thyn-ai/mojo-kernels/blob/main/README.md#fuse-mojo--fuzzy-search-fusejs-drop-in

## 4. BM25 versus an optimized scorer

Category: Retrieval
Result: 1.10–1.95× warm-query speedup across the tested grid.
Workload: M4 Max · 1K–100K docs · 5–20 query terms
Comparison: versus the best measured bm25s / Numba variant
Evidence status at snapshot: Published benchmark
Contribution: Built a compiled BM25 scorer with precomputed sparse scores and query batching.
Detail: Nine workload cells compare the best measured paths. The separate rank_bm25 comparison reports much larger 149×–13,137× ratios, but it is a different baseline and is not the headline here.
Limitations: Scoring variants differ between libraries. Numerical parity is measured against rank_bm25, not bm25s. The smaller optimized-baseline range is the headline.
Validation scope: Maximum reported absolute score difference versus rank_bm25: 3.6e-15. A limited top-10 overlap check with bm25s reports 1.00.
Evidence quality: Source-reported ratios are retained; rounded table cells should not be used to recalculate exact ratios.
Portfolio section: https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-bm25
Recorded date in source: 2026-09-20
Two-baseline comparison: https://github.com/thyn-ai/mojo-kernels/blob/main/README.md#benchmarks

## 5. Molecular wavefunction grids

Category: Kernels
Result: 72× speedup versus the vectorized baseline.
Workload: M4 Max · single thread · benzene / 6-31G* · 50³ grid
Comparison: versus the cclib pyquante2 / NumPy path
Evidence status at snapshot: Published benchmark
Contribution: Built the native Gaussian-grid evaluation path and Python integration.
Detail: The selected comparison uses the NumPy-vectorized pyquante2 backend. Reported maximum absolute difference is 8.4e-13 for this workload.
Limitations: A 50³ single-orbital grid, not an entire chemistry simulation. The much larger legacy-PyQuante ratio is intentionally not the headline.
Validation scope: Maximum reported absolute difference: 8.4e-13 against the selected vectorized baseline.
Evidence quality: Source rounds the speedup to 72×; absolute timings are rounded too.
Portfolio section: https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-scientific
Recorded date in source: 2026-09-19
Benchmark record: https://github.com/thyn-ai/mojo-kernels/blob/main/README.md#cclib-mojo--electron-density--wavefunction-grids-cclib-methodvolume
Baseline detail: https://github.com/thyn-ai/mojo-kernels/blob/main/python/cclib_mojo/README.md

## 6. Filtered vector retrieval

Category: Retrieval
Result: 4,099QPS at concurrency 8; recall reported as 1.0000.
Workload: Cohere-100K · 768 dimensions · cosine · 1% filter
Comparison: absolute throughput; no competitor-speedup claim
Evidence status at snapshot: Open upstream PR
Contribution: Authored the benchmark connector and reported the filtered-search result.
Detail: The author reports a real VectorDBBench CaseRunner against a live TCP service. The PR summary does not specify the benchmark hardware. Equality-label filtering is supported; numeric range filtering is not claimed.
Limitations: Serial p99 is reported separately as 1.5 ms. It is not p99 at concurrency 8. Connector PR is open; this is not leaderboard certification.
Validation scope: Reported recall 1.0000 and nDCG 1.0000 on the selected Cohere-100K workload.
Evidence quality: Author-reported result in an open PR; hardware is not specified in the inspected summary.
Portfolio section: https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-vector
Recorded date in source: 2026-07-21
Connector and reported results: https://github.com/zilliztech/VectorDBBench/pull/824

## 7. 27B-model concurrent serving

Category: Inference
Result: 781.3tok/s aggregate output throughput.
Workload: H100 80GB · Gemma 3 27B · FP16 · 550 in / 256 out
Comparison: absolute serving result; no cross-engine multiplier
Evidence status at snapshot: Archived result
Contribution: Built and benchmarked the serving runtime described in the source documentation.
Detail: The recorded open-loop throughput rises from 245.3 tok/s at concurrency 8, to 409.0 at 16, 585.6 at 32, and 781.3 at 50. Closed-loop batch numbers are not mixed into this result.
Limitations: Taken from the supplied August 26, 2026 documentation snapshot. The live documentation could not be rechecked for this edition.
Validation scope: Scope: 500-prompt sonnet-shaped workload, greedy decoding, prefix caching disabled, one H100 per engine.
Evidence quality: First-party archived document supplied by the user; live page and raw run were not revalidated.
Portfolio section: https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-h100
Recorded date in source: 2026-08-26
Serving documentation: https://docs.algenta.ai/benchmarks/nvidia-llm-serving

## 8. Warm deterministic execution

Category: Systems
Result: 0.83–0.93ms documented warm finance.npv execution.
Workload: One representative operation · warm resident runtime
Comparison: documented latency observation, not a speedup
Evidence status at snapshot: Documented observation
Contribution: Built the typed compute surface and managed-runtime integration.
Detail: The public documentation reports 0.83–0.93 ms for warm finance.npv. Hardware, sample count and a percentile distribution are not provided in the inspected passage.
Limitations: Excludes first-ever runtime provisioning, documented at about 110 s. Not end-to-end agent latency or a benchmark of every function.
Validation scope: This entry is deliberately classified as an observation, not a matched benchmark campaign.
Evidence quality: Public documentation reviewed; hardware and distribution are unspecified.
Portfolio section: https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-compute
Public documentation source: https://github.com/thyn-ai/sqai-mcp/blob/64e02f66768fbfb5f0da78e723e3932e86024d44/docs/faq.md
Technical docs: https://docs.sqai.com/faq

## 9. Bounded phrase-query memory

Category: Systems
Result: 19.51MB peak RSS in the selected scored test.
Workload: Repeated-token phrase stress case · scored search
Comparison: prior implementation exceeded a 256 MiB watchdog
Evidence status at snapshot: Open upstream PR
Contribution: Authored the proposed phrase-slop fix and memory-bounded matching path.
Detail: For a repeated a/b/c stress document, the author reports 19.51 MB peak RSS in scored search after the prior implementation crossed a 256 MiB watchdog. A distinct adversarial case reports 19.68 MB.
Limitations: The baseline is a watchdog lower bound, not a measured exact peak. No memory-reduction percentage is inferred. Submitted; not merged.
Validation scope: The PR reports 163,840 exhaustive three-term comparisons against brute-force span matching and 1,353 library tests passed / 8 ignored.
Evidence quality: Author-reported stress measurement in an open PR; not a general database-memory result.
Portfolio section: https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/#case-memory
Recorded date in source: 2026-11-06
Patch and stress results: https://github.com/quickwit-oss/tantivy/pull/3248

## Recorded NVIDIA serving workload cells

Qwen2.5-7B, BF16, NVIDIA A40, 64 output tokens; three repetitions per cell. Arms are forced recomputation, fresh-process persistent-cache hit, and same-process native-hot cache. Timings are recorded, not generated by the animation.

Workload | Recompute mean batch seconds | Persistent-hit mean batch seconds | Native-hot mean batch seconds
8K prefix · C=1 | 2.985708 | 2.913860 | 1.878551
8K prefix · C=4 | 6.572369 | 3.343736 | 2.029236
8K prefix · C=8 | 11.383825 | 3.716415 | 2.146576
16K prefix · C=8 | 23.655112 | 5.396244 | 2.418234

## Contact

Fully remote scoped engineering projects: https://headless-ynhohfjncyy-angelmarinov1-140d.wix-site-host.com/contact.html