Tempe, AZ · Open to 2026 internships & nationwide relocation
I build the layer
between research
and the metal.
ML systems engineer working where model architecture meets hardware — hand-written
CUDA kernels, 4-bit quantized LLMs that fit on a laptop,
multi-agent systems that reason across sessions, and the
MLOps pipelines that keep all of it honest in production.
~175 tok/s MLX decode on Apple Silicon◆1.7× speedup on fused CUDA LayerNorm◆0.9998 cosine parity vs CUDA baseline◆96.4% brain-tumour classification accuracy◆0.74% MAPE — SOTA hybrid-energy forecast◆<6 GB VRAM for a 3B fine-tuned LLM◆<30 ms production inference latency◆2 peer-reviewed Elsevier publications◆
~175 tok/s MLX decode on Apple Silicon◆1.7× speedup on fused CUDA LayerNorm◆0.9998 cosine parity vs CUDA baseline◆96.4% brain-tumour classification accuracy◆0.74% MAPE — SOTA hybrid-energy forecast◆<6 GB VRAM for a 3B fine-tuned LLM◆<30 ms production inference latency◆2 peer-reviewed Elsevier publications◆
01 Profile
Most people stop at "the model works." I'm interested in what happens after.
I'm Vigneshkumar Venugopal — an MS Data Science candidate at
Arizona State University and a former research engineer at
ERNET India, where I shipped medical imaging pipelines that radiologists actually used.
My work lives in the unglamorous middle: the fused kernel that makes training 1.7× faster, the
quantization scheme that drops a 3B model under 6 GB of VRAM, the memory architecture that stops
an agent from contradicting itself three sessions later. I write CUDA when PyTorch is too slow,
port models to MLX when CUDA isn't there, and wrap the result in FastAPI, Docker and CI so it
survives contact with reality.
Alongside that, I've co-authored two peer-reviewed papers in Elsevier journals on
deep learning for renewable-energy systems — a Transformer-LSTM ensemble that beat every RNN
baseline we tested, and a multi-objective NSGA-II optimizer for green hydrogen production.
Currently looking for ML engineering / ML systems roles and internships.
02 Selected work
Ten projects, one thesis: make big models run on small machines.
Every project below is open source. Click a card to read the code.
01
QLoRAQwen2.5-3Bllama.cppGGUF
NeuralScholar
A 3B language model fine-tuned to research and explain ML/DL concepts — locally.
Fine-tuned Qwen2.5-3B with QLoRA (4-bit NF4 + double quantization) on a 100K-sample corpus
I built myself from ArXiv, StackExchange and Distill.pub — with MinHash LSH deduplication stripping ~35% of
redundant data. A two-stage router pairs the local SLM with frontier models (Gemini / Claude), so the
cheap model handles what it can and only escalates when it must.
A GPT-2 class transformer written from scratch — then made fast by hand.
Decoder-only transformer built from zero in PyTorch and trained on TinyStories, then I went a layer
deeper: three progressive versions of a fused LayerNorm CUDA kernel using shared-memory
tree reduction, Welford's algorithm with warp shuffles, and vectorized float4 access —
profiled with Nsight Compute and NVTX markers.
A 3.3B DeepSeek-OCR model ported to Apple Silicon with zero PyTorch at runtime.
Pure MLX inference pipeline: a stacked MoE layer using batched gather_mm,
a sliding KV cache with pinned image-prompt context, and bit-exact position-embedding resizing that
matches PyTorch's bicubic-antialias interpolation via cached separable matrices.
Biologically-inspired hierarchical memory for LLM coding agents.
Coding agents forget across sessions, and worse, they don't notice when new information contradicts
what they already "know." Engram fixes both with an episodic buffer → LLM-driven consolidation
(promote / compress / discard) → long-term semantic store architecture, plus a background agent
that actively hunts for contradictions. I also wrote LHMBench, an original 6-scenario
benchmark, to prove it — it's the only system tested that detected contradictions at all.
Reads any ArXiv ML paper, emits a validated PyTorch scaffold and CUDA blueprint.
A cascading, stateful agent network orchestrating four frontier models — Gemini 2.5
Flash/Pro, Qwen2.5-Coder 32B and Nemotron Super 49B — behind scope-validation layers. Papers become a
JSON ResearchContract, which becomes model.py, train.py,
dataset.py and annotated CUDA C++ stubs. Streams over SSE.
Session-isolated RAG that builds technical memory on demand.
A local LLM extracts keywords, dual-strategy ranking queries ArXiv, and ~25 papers vectorize into
Qdrant in under 15 seconds. The interesting part is the layout-aware PDF chunker:
PyMuPDF font-size baselines split text along real semantic section boundaries instead of arbitrary
character counts.
An end-to-end MLOps pipeline for churn prediction — the boring parts done properly.
Bayesian hyperparameter search across LightGBM and XGBoost, SHAP attributions for stakeholders, and a
GitHub Actions pipeline that runs schema validation, unit tests, retraining and artifact versioning on
every commit. Evidently AI + KS-tests watch for covariate drift in production.
A distributed agent mesh for real-time flood monitoring and emergency triage.
Built on Google's Agent Development Kit: specialized roles — FloodMonitor, TriageAgent,
Medical and Infra Sentinels — coordinated through asyncio for non-blocking parallel
situational analysis, with separated modules for agent lifecycles, tool integration and consolidated
mission reporting.
Predicting CVE severity from raw text — fine-tuned encoder vs. zero-shot LLM.
A comparative triage engine that pits zero-shot LLM inference against a fine-tuned cybersecurity
encoder (SecBERT) for mapping natural-language descriptions to CVSS v3.1 severity, then an
agentic layer turns raw model output into analyst-ready triage reports.
Autonomous multi-agent code review and self-healing.
Built on NVIDIA's NeMo Agent Toolkit — agents review a codebase, identify defects, and propose or
apply their own fixes rather than just filing complaints.
Architected a multi-class classification and segmentation pipeline identifying glioma, meningioma and pituitary tumours, reaching 96.4% accuracy on real-world MRI datasets.
Developed a ResUNet semantic segmentation model at a 0.89 Dice coefficient, automating tumour boundary delineation and cutting manual volume-estimation time for radiologists by 70%.
Optimized high-resolution NIfTI / DICOM data-loading with MONAI — a 35% reduction in training latency and efficient multi-GPU training.
PyTorchMONAIResUNetDICOMMulti-GPU
04 Publications
Peer-reviewed work in Elsevier journals.
Deep learning and evolutionary optimization applied to renewable energy systems.
Renewable Energy · 2025
10.1016/j.renene.2025.122369
A novel ensemble machine learning approach for optimizing sustainability and green hydrogen production in hybrid renewable-based organic Rankine cycle-operated PEM electrolyser systems
Designed an ensembled Transformer-LSTM architecture, fusing multi-head attention with LSTM layers to capture both long- and short-range temporal dependencies and sidestep the vanishing-gradient and hyperparameter sensitivity of plain RNNs.
Applied integrated gradients for interpretability, confirming the network's attributions aligned with real-world physical constraints — with 5-fold cross-validation for generalizability.
Achieved a 0.74% MAPE, outperforming LSTM (0.89%), RNN (1.69%) and FFNN (2.1%) baselines across multi-variable time-series forecasting.
Analysis and multi-objective evolutionary optimization of a solar-biogas hybrid cascade Kalina organic Rankine cycle for sustainable cooling and green hydrogen production
Built simulation models in Python with CoolProp, automating dynamic thermodynamic property acquisition and continuous 4E (energetic, exergetic, economic, environmental) evaluation under variable boundary conditions.
Architected an NSGA-II multi-objective optimization framework across 400 generations, maximizing energy utilization and exergy efficiency while minimizing total system cost — producing a full Pareto frontier.
Implemented a two-stage MCDM pipeline — CRITIC for objective weighting, then TOPSIS and MABAC for distance-based ranking — to select the optimal configuration from high-dimensional Pareto data.
SQLPostgreSQLMySQLMongoDBCockroachDBRedisKafkaCelerySparkHadoop MapReduceCassandraFAISSQdrantChromaMilvusPineconeElasticsearchGCPVertex AI
Languages & Foundations
PythonC++CJavaBash / ShellJuliaMATLABNumPyPandasMatplotlibSeabornOpenCVPILLinuxStatistical ModelingExploratory Data Analysis
06 Contact
Building something that needs to run fast and stay up?
I'm open to ML engineering and ML systems roles, research collaborations, and 2026 internships —
based in Tempe, AZ and happy to relocate anywhere in the US.