DHRUV GARG

DHRUV GARG

AI / ML ENGINEER

I ship PyTorch models to production — computer vision, generative systems, and the full stack around them.

Stratum I · 1200m

SP-01 · Searchlight Protocol

Coarse-to-Fine
Detection

Small-object detection in high-resolution aerial imagery. A ResNet18 + LayerCAM guide fuses backbone layers 2/3/4 into a semantic heatmap; an NMS slicer extracts candidate crops; a YOLOv8 detector runs only on those — skipping 80%+ of empty background before fine inference.

PythonPyTorchResNet18LayerCAMYOLOv8OpenCV
Searchlight architecture
coarse detector → ROI slice → fine detect
Stratum II · 2400m
Neural Canvas architecture
VGG-16 perceptual loss → transformer → stylized output

NC-03 · Neural Canvas

Perceptual
Style Transfer

Feed-forward style transfer (Johnson 2016): 5 residual blocks with Instance Normalization, trained against a VGG-16 perceptual loss on MS-COCO across 4 styles. Shipped to a live Hugging Face Space with a callable /stylize API and ONNX export.

PyTorchVGG-16InstanceNormMS-COCOONNX
Stratum III · 3600m

PQ-02 · PixelQueue

Async Annotation
Infrastructure

Human-in-the-loop vision pipeline: a React-Konva canvas behind an NGINX gateway → FastAPI → Celery workers, with PostgreSQL, a Redis broker, and MinIO blob storage. YOLO/SAM auto-labeling workers, COCO + YOLO export, an SSE event stream, and RBAC across admin / reviewer / annotator roles.

FastAPICeleryRedisPostgreSQLMinIOYOLO/SAM
car: 0.94
React-Konva
POST /api/v1/annotate
{"id": 1042, "class": "car"}
202 Accepted
FastAPI
#1041OK
#1042RUN
#1043
Celery
annotations
1042car0.94
1041truck0.89
PostgreSQL
PixelQueue architecture
React UI → FastAPI → Celery → auto-label / export workers → PostgreSQL
Stratum IV · 4800m
pygog — agentic CLI

$ pygog "Schedule a meeting with the ML team for Thursday"

Parsing intent…

Detected: Calendar.CreateEvent

Resolved: ML Team → 5 contacts

Found slot: Thu 2:00 PM – 3:00 PM

Event created · 5 attendees notified

$

PyGOG CLI architecture
Typer CLI · 5-provider LLM intent parsing · OAuth2 tool routing

PG-04 · PyGOG CLI

Agentic
Workspace CLI

Natural language in, tool routing out. A Go→Python reimplementation of gogcli that adds an LLM agent layer over 5 providers (DeepSeek, OpenAI, Gemini, Anthropic, OpenRouter), orchestrating across Gmail, Drive, Calendar, and Tasks — with web search, OS-keyring credential storage, and multi-account aliases.

TyperLLM Agents5 ProvidersOAuth2Keyring
Core Sample · 6000m

Journey

2022 Foundation

Deep Learning

CNN architectures, training dynamics, evaluation pipelines from scratch.

PyTorchNumPyMatplotlib
2023 Vision

Computer Vision

Monocular depth, custom detectors, real-time inference pipelines.

YOLOv8OpenCVLayerCAM
2023 Creative

Generative AI

GANs, neural style transfer, VGG perceptual loss research.

GANsNSTVGG-16
2024 Product

Shipped to Prod

Neural Canvas on Hugging Face — research notebook to ONNX export.

ONNXHuggingFaceDocker
Now Systems

Full-Stack ML

Async infra, LLM agent orchestration, production deployment.

FastAPICeleryLLMs
Sensor Matrix · 7200m

Capabilities

Every capability below is bound to a system I shipped — not a self-rating.

  • 01

    Perception

    Proven in Searchlight Protocol →

    Coarse-to-fine detection of small objects in high-resolution aerial imagery — slicing regions of interest and skipping 80%+ of empty background before fine inference.

    • PyTorch
    • YOLOv8
    • LayerCAM
    • OpenCV
    • ONNX
  • 02

    Generation

    Proven in Neural Canvas →

    Feed-forward neural style transfer trained against a VGG-16 perceptual loss, exported to ONNX and shipped as a live demo on Hugging Face.

    • PyTorch
    • VGG-16
    • ONNX
    • Hugging Face
    • GANs
  • 03

    Infrastructure

    Proven in PixelQueue →

    Async, human-in-the-loop annotation at scale: Celery workers behind a FastAPI gateway, a Redis broker, a PostgreSQL store, and a React-Konva labeling canvas.

    • FastAPI
    • Celery
    • Redis
    • PostgreSQL
    • Docker
  • 04

    Agents

    Proven in PyGOG CLI →

    Natural-language intent parsed into authenticated tool calls over 5 interchangeable LLM providers, orchestrated across Google Workspace — Gmail, Drive, Calendar, and Tasks — from a single command line.

    • Typer
    • LLM Agents
    • OAuth2
    • Function Calling
Verification · 8400m

Proof

+
Open Source Contributions
+
LeetCode Questions Solved
+
LC Contest Rating
250+ open source contributions, 550+ questions solved on LeetCode, 1760+ LeetCode contest rating
© 2026 Dhruv Garg

Open to AI/ML research, internships,
and full-stack roles where perception systems matter.