Inside the work
5.11×
faster autoregressive
generation
GENERATION LATENCYLOWER IS BETTER ↓
Measured benchmark · Autoregressive generation
GPU Inference Optimization LabView on GitHub (opens in a new tab)
AI / ML ENGINEER
Open to opportunitiesGollan
Kurulkar.
Better systems.
Backed by evidence.
I build, profile, and improve LLM systems. From GPU inference to evaluation pipelines, I turn bottlenecks into measurable progress.
01 / Selected work
The work.
The evidence.
Real systems. Controlled experiments.
The gains, the tradeoffs, and what I learned.
LLM inference · 2026/01
+49%output throughput
OUTPUT THROUGHPUTHIGHER IS BETTER ↑
Measured benchmark · Latency tradeoffs documented
Quantized LLM Serving
A reproducible FP8 serving benchmark that cut model weights by 42.8% and increased output throughput from 4,805 to 7,159 tokens per second, with latency tradeoffs reported alongside the gain.
View on GitHub (opens in a new tab)GPU performance · 2026/02
5.11×faster generation
GENERATION LATENCYLOWER IS BETTER ↓
Measured benchmark · Autoregressive generation
GPU Inference Optimization Lab
An optimization study of PyTorch inference paths. Autoregressive generation improved from 552.6 ms to 108.2 ms, while manual CUDA Graph capture reached 4.30× faster per-step execution.
View on GitHub (opens in a new tab)More experiments & systems05 projects
- 03Distributed training · 2026
DDP Scaling on H100
A GPT-2 Large scaling study that explains why four H100 nodes ran 2.4× slower than one: 93.5% of the representative step was synchronization over an NCCL socket fallback without RDMA.
View on GitHub0.41×4-GPU speedupView repository (opens in a new tab) - 04Reliability engineering · 2026
LLM Inference Observability
A load-testing and telemetry project that reduced P95 latency from 98.5 seconds to 3.44 seconds, eliminated observed timeouts, and raised request success from 32% to 87%.
View on GitHub98.5→3.44sP95 latencyView repository (opens in a new tab) - 05MLOps · 2026
Coding-Agent Evaluation Pipeline
An Airflow- and MLflow-based evaluation workflow for coding agents, with isolated execution, artifact tracking, and a tiny three-task SWE-bench Verified run used to validate the pipeline—not claim a model benchmark.
View on GitHub2 / 3pipeline validation runView repository (opens in a new tab) - 06Retrieval evaluation · 2025–2026
Financial Document RAG Evaluation
An evaluation-first RAG study across 42 SEC filings and 100 FinanceBench questions, comparing retrieval, reranking, and prompt policy while improving answer correctness from 0.23 to 0.34.
View on GitHub+48%answer correctnessView repository (opens in a new tab) - 07Negative-results study · 2026
Level-Adaptive QLoRA
A publishable record of a constrained fine-tuning experiment that did not establish the intended result. The repository documents the CPU-only fallback, lack of 4-bit loading, 60-step run, and limits on interpretation.
View on GitHub60steps, limits documentedView repository (opens in a new tab)
02 / Expertise
From the bottleneck
to the bigger picture.
I connect low-level performance with the reliability of the whole system.
- 01
GPU & LLM performance
Benchmarking and optimization across model serving, autoregressive decoding, distributed training, quantization, and GPU execution.
PyTorch · CUDA Graphs · vLLM · H100 · NCCL - 02
Evaluation & reliability
Controlled experiments that expose bottlenecks, quantify tradeoffs, and keep negative or inconclusive results visible.
MLflow · RAGAS · SWE-bench · Locust · RCA - 03
Applied LLM systems
Retrieval, agent, and inference services designed around measurable quality, typed interfaces, and reproducible behavior.
LangGraph · FastAPI · FAISS · FastMCP · Pydantic - 04
MLOps & experiment systems
Repeatable evaluation workflows with orchestration, experiment tracking, containers, tests, and CI-ready project structure.
Airflow · MLflow · Docker · GitHub Actions · pytest
How I work01 Profile the bottleneck02 Change one lever03 Verify the result
03 / A little about me
An engineer’s
instinct.
A systems mindset.
Before profiling models, I was investigating physical processes. The question has always been the same: what is the system actually doing?
My path runs from electronics engineering and manufacturing into data, backend software, and ML systems. I bring that hands-on perspective to building practical tools and making messy information understandable.
More on LinkedIn (opens in a new tab)The path so far
- 2024 — PresentCurrent
Process Engineer
UNI — Sophisticated Electronic Assembling
Investigate production and quality problems across PCB assembly from inspection data, using root-cause analysis and process knowledge alongside production, quality, and engineering teams.
- 2023
Data Analyst
Eleos Health
Built an automated PDF-to-data pipeline, analyzed PostgreSQL datasets, and enabled A/B testing that measured a 53% improvement in documentation speed.
- 2020 — 2022
Data Analyst / Developer
Freelance
Delivered Python data-collection and validation pipelines for software-development clients, producing reusable JSON and CSV datasets.
Have something in mind?
Let’s make it work.Let’s build
what’s next.
Start a conversation gollankurulkar@gmail.com
ML systems, evaluation challenges,
or the next great team.