Anyone can claim GPU experience. Here's what you can click on:

Most of it is unglamorous. The Dockerfile that builds on the first try. The installer that survives a Vast.ai box shipping the wrong CUDA driver. The health check that had been quietly dropping half your services since March and nobody noticed.

I came up through research before infrastructure, which mostly means I can read a paper on Monday and have it deployed by Friday instead of bookmarking it and feeling good about myself.

People hire me to:

• Deploy models as serverless GPU endpoints on RunPod, Modal, or Vast.ai. Dockerized, CI/CD, autoscaled down to near-zero idle cost.
• Build RAG and document pipelines that can prove they work: retrieval, citation tracking, multi-hop reasoning, LLM-as-judge evals so "it feels better" becomes a number.
• Stand up multi-agent systems, including latent-space designs where agents pass hidden states instead of burning tokens regenerating text.
• Serve hot-swappable LoRA adapters off one base model, routed by domain. I train the adapters too.
• Rescue deployments that almost work. Multi-GPU OOM, driver mismatch, topology-aware placement, squeezing a model onto a card that shouldn't fit it.

ODS assumed every service spoke HTTP. Anything exposing only TCP or a CLI failed its health check, got marked dead, and disappeared from the dashboard without an error anyone would see. Silent data loss, basically, but for services.

I threaded a health_type: http|tcp|none field through the whole stack: JSON schema, catalog generator, dashboard API, and the shell scripts underneath. The interesting part wasn't the feature, it was the migration. Dozens of existing manifests had to keep working untouched, so the default had to infer correctly from what was already declared.

Same codebase, also merged: a node diagnostics endpoint for runtime hardware introspection, gpu_backends schema validation, and compatibility blocks across 25 manifests.

Multi-agent systems waste enormous money re-serializing thoughts into English so the next agent can re-parse them. LatentMAS (ICML 2026 Spotlight) showed you can skip that and pass hidden states directly.

I extended it with role-specialized LoRA adapters that hot-swap at runtime, so one base model plays Planner, Critic, Refiner, and Judger without four sets of weights in VRAM. Critic and Refiner never generate a token, about 200ms each. Ships with VLM support on Qwen2.5-VL-7B, RAG integration, and a RunPod serverless deployment with CI/CD.

The authors added it to their README as community extension #5, next to work from MIT's LAMM group. Demo video.

The README is also explicit about what it isn't: PEFT-based adapter management, not true S-LoRA. No custom CUDA paging kernels, no heterogeneous batching. I'd rather you read that there than discover it on a call.

Vast.ai and similar marketplaces hand you a box that is technically the GPU you paid for and broken in a dozen quiet ways. Wrong driver, missing ACLs, a container runtime that isn't wired up.

I wrote a multi-phase hardened installer for it: set -euo pipefail throughout, hard-fail ACL checks, GPU-tier detection, and roughly 28 documented host-environment failure modes with the fix for each. That documentation was most of the value. Anyone can write the happy path.

I care where this ends up. My research has been in solar cell optimization and medical imaging, and the client work I chase is in health, energy, education, and scientific tooling. Good rates matter, I'm not shy about that. But I do the best work when the thing being built has some reason to exist beyond a funding round.

If that's what you're building, lead with it.

GPU & serving — RunPod, Modal, Vast.ai, vLLM, Docker, Docker Compose, torch.distributed, FSDP, 4-bit quantization

LLM & agents — PyTorch, Transformers, PEFT/LoRA, LangChain, MCP, RAG, LLM-as-judge evaluation, Qwen2.5-VL

Backend & ops — Python, FastAPI, Bash, GitHub Actions, pytest, ShellCheck, SQL, MongoDB

Also — Diffusers, TensorFlow, JavaScript, TypeScript, Git

5 peer-reviewed papers · 75 citations · h-index 5 — Scholar · ORCID · ResearchGate

_{Quartiles follow the latest JCR/Scopus release and vary by database and subject category.}

B.Sc. Electrical & Electronic Engineering, University of Rajshahi (AI & Solar Energy Laboratory). SQL Advanced and REST API Intermediate (HackerRank), Deep Learning with TensorFlow (IBM), Prompt Engineering (Vanderbilt).