JG
SYSTEMS // UCSD 2028
Connect
UC SAN DIEGO · CLASS OF 2028 LA JOLLA, CA · SYSTEMS × CUDA × BIOINFORMATICS

JAIDEEP
GORIJAVOLU

Building AVX2 / CUDA exact genomic aligners, lock-free C++20 vector search engines, and statistically rigorous ML evaluation harnesses. Porting TILE to CUDA & targeting Summer 2027 SWE / Systems roles.

View Selected Work GITHUB ↗ LINKEDIN ↗
22–23× MM
TILE D71 · 2,472 Exact
6,488 QPS
vecsearch · 0.23ms p99
50.4 GCUPS
SSE2 Smith-Waterman
0.950 F1
SimCLR Sensor Probe
SCROLL TO SPLIT DNA INTO PROJECT CARDS · DRAG TO ORBIT FULL 3D WORLD EXPLORE WORK ↓
STAGE 02 // HELIX_UNZIPPED → ARCHITECTURE MODULES

Selected Systems & AI

Click any module to morph the full-screen 3D topology & inspect

LIVE GITHUB TELEMETRY
Loading recent pushes…
STAGE 03 // FULL_SCREEN_NEURAL_NET · EXPERIENCE

Engineering & Research Log

Click any node to expand verified telemetry

  • Built catbench (uv run catbench run): 25 queries (19 test, 6 dev), 6 multi-step workflows, ~1,990 graded pairs across 269 catalog listings with 95% bootstrap CIs.
  • Test split proved text-embedding-3-small beats BM25 across all metrics (R@10 0.801 vs 0.651, nDCG@10 0.776 vs 0.631, paired Δ +0.145, MRR 0.921 vs 0.733).
  • Benchmarked 55 configurations (4 embedding models + BM25 hybrid; 243 tests; CEO reproduced all 55 runs) and proposed a one-line provider-pinning fix for hosting latency variance.
  • Built TILE (C++ / AVX2 SIMD): stores compressed 2-byte tile borders and recomputes only optimal-path tiles, cutting traceback memory from O(mn) to O(mn/T + T²).
  • D63 benchmark (Xeon Silver 4216, 1 core): exact score agreement across all 2,472 pairs (ONT 1k–500k, SARS-CoV-2, PREFAB-R, BAliBASE 3); beats Myers–Miller by up to 25.4× on ONT-500k (68.0 s vs 1,728 s; 7.0× faster on 8 threads).
  • Latest D71 (Steps 14–18): dna_twilight_like runs 22–23× faster than Myers–Miller (up from 5×); porting to CUDA for TWILIGHT.
  • Built automated Python pipelines validating HL7/FHIR R4 clinical schemas (DiagnosticReport, Patient) and pytest/Pydantic mocks stress-testing EHR APIs.
  • Built evaluation harnesses benchmarking clinical LLM responses on hallucination rate, latency, and semantic similarity across prompts, plus embedding retrieval recall analysis.

Authored computational software engineering evaluation tasks; 3 tasks accepted after expert peer review.

Conducted Quantitative Phase Imaging (QPI) biophysical studies measuring elasticity and morphology of mammalian cells.

Systems & Perf
C++17/20 · CUDA · AVX2/SSE2/NEON · CMake · pybind11 · perf · Google Benchmark · ASan/UBSan/TSan
ML & IR Eval
PyTorch · HF Transformers · LoRA · SimCLR · Faiss · HNSW/ANN · BM25 · nDCG / MRR · Bootstrap CIs
Backend & Infra
Python · FastAPI · Pydantic · SQLAlchemy · SQL · HL7/FHIR R4 · Docker · Linux · pytest / CTest
STAGE 04 // HARDWARE & IR TELEMETRY LAB

Verified Benchmarks

STAGE 05 // CONTACT

SUMMER 2027
LET'S BUILD.

Targeting software engineering, C++/CUDA systems, and applied ML/IR internships. Open to discussing low-level systems optimization, computational genomics, and ML evaluation infrastructure.

Send Email → GitHub (7 Repos) ↗ LinkedIn ↗
COPIED jgorijavolu@ucsd.edu