★ SHIVAM STUDIOS PRESENTS ★
VOL. 1·ISSUE #001·$3.99
★  ALL-NEW!  ORIGIN ISSUE  ★
SHIVAM
SINGH
AI / ML ENGINEER  •  FULL-STACK DEVELOPER
PYTHONPYTORCHLANGCHAINNEXT.JSTYPESCRIPTAWS
“THE MOST EXCITING ML ENGINEER IN PRINT”★★★★★
PG 2|CHAPTER ONE: ORIGIN

CHAPTER 1: THE ORIGIN

SOMEWHERE IN SANTA BARBARA… A HERO EMERGES
SHIVAM
SINGH

“I don't just build models — I engineer intelligent systems that solve real problems.”

AI/ML engineer & full-stack developer building intelligent systems.
THE BEGINNING…

CHAPTER ONE

It started at UC Santa Barbara — a Statistics & Data Science major who fell in love with the math behind machine learning and never looked back.

LEVEL UP!

THE TRAINING ARC

ML research at two UCSB labs. An AI/ML internship at MaiHealth. NeurIPS reviewer. Multi-agent systems at 2 am. Every challenge forged a sharper mind and a deeper stack.

→

NOW ARMED WITH THE FULL STACK —  PYTORCH • LANGCHAIN • NEXT.JS • TYPESCRIPT • PYTHON • AWS

→
PG 3|CHAPTER TWO: PROJECT FILES

CHAPTER 2: PROJECT FILES

FEATURED STORY — ISSUE #1
TRIBEMAP

Brain Response Prediction

Predicts how the human brain responds to any image, video, or text using Meta's TRIBE v2 model. 20,484 cortical vertices visualized in real time via Three.js, powered by a trimodal encoding pipeline (V-JEPA2, Wav2Vec-BERT, LLaMA 3.2) on Modal GPUs with Claude-generated neural interpretations.

TRIBE v2Three.jsModalClaudeVIEW PROJECT →
BATTLE ANGEL

Humanoid disaster-rescue RL agent — PPO policy trained across 8 MuJoCo gym environments on LIDAR, cameras, thermal & accelerometer inputs. Gemini 3.5 Flash generates unseen scenarios on the fly; live Three.js visualization. Built in 6 hours.

PPOMuJoCoGeminiThree.js
DRISHTI

Vision-language chest X-ray screening for clinics without reliable bandwidth. Qwen2-VL fine-tuned on TBX11K with Grad-CAM heatmaps and an offline GGUF path for on-device inference.

Qwen2-VLGrad-CAMGGUFModal
SCARTISSUE

Indexes a repo's git history of fixes and warns when a PR is about to reintroduce a regression your team already paid for. MCP server + web UI cross-reference every hunk against known scar patterns.

MCPGitHubAnthropicPython
CONNECTAI — VR CHARACTERS

RAG-based AI character system for a no-code VR platform — Mistral + ChromaDB NPCs with persistent episodic memory and real-time reasoning via JSON-based Unity integration.

MistralChromaDBRAGUnity
NEURIPS VISUAL AGENTS

Evaluation infrastructure for NeurIPS 2025 competition benchmarking 50+ visual agents against biological baselines using PyTorch and Weights & Biases.

PyTorchW&BNeurIPS
MAIHEALTH PIPELINE

Voice-to-FHIR pipeline achieving 94% extraction accuracy across 15+ medical entity types. Plus a RAG chatbot with LangChain & ChromaDB querying EHR data with <2s latency.

LangChainChromaDBFastAPI
NEMOLENS

Multimodal video analysis pipeline using NVIDIA Nemotron Nano 12B & local Whisper — parallel frame extraction, structured notes, quizzes & a voice assistant. Submitted to NVIDIA GTC 2026.

NVIDIAWhisperStreamlitPython
POWERTRACKER

Full-stack job tracker with AI-driven insights — Next.js 15, TypeScript, Tailwind, Postgres/Prisma, deployed on Vercel with 8 active users.

Next.jsTypeScriptPrismaAnthropic
MULTI-AGENT LLM TRADING

End-to-end multi-agent crypto trading system using LangGraph — 8 specialized LLM agents (GPT-4, Claude) via a shared state graph with production-grade data pipelines from YFinance, CoinGecko, Reddit & Binance.

LangGraphGPT-4ClaudePython
PG 4|CHAPTER THREE: INTEL FILES

CHAPTER 3: INTEL FILES

ML PAPERS — DECODED & DISPATCHED

INTEL DISPATCH — ISSUE #1
READ ON MEDIUM ↗
WHY AI SHOULD THINK IN CONCEPTS, NOT WORDS

Diffusion Language Models Explained

The future of language AI isn't next-token prediction. Explore how diffusion-based models that reason in concept space outperform autoregressive architectures — and why this changes everything about how we build LLMs.

Diffusion LMConcept SpaceNLPREAD FULL BRIEF →
ISSUE #2 — DEEP LEARNING
RESNET UNPACKED

How Skip Connections Unlocked Deep AI

How residual learning solved the vanishing gradient problem and made 100+ layer networks trainable — the architecture still powering modern vision models.

ResNetSkip ConnectionsVision
ISSUE #3 — CLASSIC ARCHITECTURES
UNDERSTANDING VGGNET

Simplicity, Depth & a Lasting Legacy

How Oxford's VGGNet proved that stacking small 3×3 filters deep beats large kernels — shaping the blueprint for every CNN that followed.

VGGNetCNNsComputer Vision
ISSUE #4 — CLASSIC ARCHITECTURES
UNDERSTANDING ALEXNET

The Shot Heard Round the AI World

The 2012 ImageNet breakthrough that ignited the deep learning era — ReLU, dropout, GPU training, and the architecture that started it all.

AlexNetImageNetGPU Training
ALL ARTICLES PUBLISHED ON MEDIUM · SHIVAM SINGHVIEW ALL →
PG 5|CHAPTER FOUR: POWER STATS

CHAPTER 4: POWER STATS

CHARACTER PROFILE — CLASSIFIED
SS
SHIVAM SINGH
CLASS: AI/ML ENGINEER
LEVELMAX
EDUCATIONUCSB '26
ALIGNMENTCHAOTIC GOOD
ROLE TYPERESEARCHER & BUILDER

SPECIAL MOVES

▶
AGENT ARCHITECTMulti-agent LLM systems from scratch
▶
RAG MASTERYSub-2s retrieval, 95%+ relevance
▶
PIPELINE FORGEEnd-to-end ML pipelines, production-grade
▶
FULL-STACK DEPLOYIdea → shipped app with real users
POWER LEVEL ANALYSIS — TOP SECRET

ML & AI

PYTHON95%
PYTORCH / TENSORFLOW88%
LANGCHAIN / LANGGRAPH90%

FRAMEWORKS & DATA

NEXT.JS / TYPESCRIPT85%
SQL (POSTGRES / SQLITE)82%
RAG / VECTOR DATABASES88%

INFRASTRUCTURE

AWS (LAMBDA, EC2, S3)75%
DOCKER / GIT / DEVOPS80%
FASTAPI / FLASK / STREAMLIT85%
THESE LEVELS ARE ALWAYS UPGRADING. LAST CALIBRATED: TODAY.
PG 6|EPILOGUE: SIGN OFF

EPILOGUE: SIGN OFF

★  ISSUE #001 — FIN  ★
TO BE
CONTINUED…
NEXT ISSUE:  THE NEXT BIG PROJECT
LETTERS TO THE HERO

Got a project? Want to build something incredible together? The hero is accepting new missions.

© 2026 SHIVAM SINGH — ALL RIGHTS RESERVED

NO HEROES WERE HARMED IN THE MAKING OF THIS PORTFOLIO