SK.

SAKET KUMAR

Data Scientist · AI Engineer

An interactive 3D journey through the world I'm building

scroll to travel · best experienced fullscreen

SCROLL TO EXPLORE
Saket Kumar

DATA SCIENTIST · AI ENGINEER · ML ENGINEER

SAKET
KUMAR

Turning data into intelligence. Building AI for real-world impact.

15+Projects Built
BScDS & AI (Hons.) · IIT Guwahati
🤖 AI / Deep Learning 🎓 IIT Guwahati 📍 Delhi, India 🟢 Open to Work

// WHO AM I

About Me

Saket Kumar

I'm Saket Kumar — a Data Scientist & AI Engineer pursuing my BSc (Hons.) in Data Science & AI at IIT Guwahati. I build AI that ships — not just demos, but systems that are trained, tested and deployed live.

My playground spans LLM agents & RAG, computer vision, on-device speech models and satellite imagery. My strength is going end-to-end: model → evaluation → API → UI → production, with measurable results — F1, ANLS, WER — backing every claim.

  • 📍 LocationDelhi, India
  • 🎓 DegreeBSc (Hons.) DS&AI — IIT Guwahati
  • 📧 Emailk.saket@op.iitg.ac.in
  • ⚡ StatusOpen to internships & collaborations

// WHAT I KNOW

Technical Skills

💻 Languages

PythonTypeScriptJavaScriptSQLRC

🧠 ML & Deep Learning

PyTorchScikit-learnXGBoostOpenCVONNX RuntimeWebAssembly

💬 Generative AI & LLMs

HuggingFaceRAGBM25 RetrievalNLIWhispersherpa-onnx

🛠️ Full-Stack & Cloud

ReactNext.jsFastAPIThree.jsNode.jsGitHub ActionsVercelDocker
🤖

// PROJECT CATEGORY 01

LLM & Agentic AI

Agents, RAG and evaluation — systems that reason, retrieve and self-correct.

5 projects

// WHAT I'VE BUILT

01 LLM & Agentic AI

RepoLens

LLM code-review agent with repo-level context — reads a PR diff plus the whole repository via BM25 + reference-graph retrieval, then posts line-level review comments to GitHub. 65 tests, CI green.

PythonLLM AgentRAGGitHub Actions
02 LLM & Agentic AI

FaithGuard

Hallucination detection & mitigation for RAG — claim-level NLI checks whether each answer is supported by its retrieved context, then auto-corrects via re-retrieval. F1 = 0.887 on SQuAD-based eval.

PythonNLIRAGLLM Eval
03 LLM & Agentic AI

DocVQA

Visual question answering over documents — OCR + layout analysis + BM25 retrieval, answered by a grounded LLM. Scores ANLS 1.000 on a 27-question eval, deployed serverless.

PythonOCRRetrievalFastAPI
04 LLM & Agentic AI

Multi-Agent Research Copilot

A team of cooperating AI agents — planner → searcher → writer → critic — that research, cross-verify and synthesize complex topics into cited survey reports with quality evaluation.

PythonAgentic AIOrchestrationLLMs
05 LLM & Agentic AI

Medical Report Summarizer

Clinical NLP that summarizes radiology & pathology reports for patients and flags urgency — deterministic triage core with a safety-gated LLM layer. 100% on a 47-case safety eval, deployed live.

Clinical NLPLLMSafety Eval
👁️

// PROJECT CATEGORY 02

Computer Vision & Remote Sensing

From deepfake forensics to satellite flood mapping — vision that sees what matters.

2 projects
06 Computer Vision

Deepfake Detection

Spatial-temporal CNN/Transformer that flags manipulated video — analyzing which artifacts generalize across generators. Research-grade benchmarking on FF++ and Celeb-DF v2.

CNNTransformerForensicsPyTorch
07 Computer Vision

Yamuna Flood Mapper

Sentinel-1 SAR satellite flood mapping for the Yamuna corridor, Delhi — free satellite data and free hosting turned into an early-warning flood layer for the city I live in.

Sentinel-1Remote SensingGIS
🎙️

// PROJECT CATEGORY 03

Speech & Audio AI

Emotion, transcription and Indic speech — audio intelligence on-device and in the browser.

3 projects
08 Speech & Audio

Svara

Privacy-first on-device voice emotion screener — 6-class speech emotion recognition (CREMA-D) running a 287K-param int8 CNN fully in-browser via WebAssembly. No audio ever leaves the device.

ONNXWASMSEROn-device
09 Speech & Audio

Meeting Intelligence

Speaker diarization → transcription → auto summary + action items + who-owns-what. In-browser ML with sherpa-onnx WASM + Whisper; WER 7.8%, action-item F1 0.947.

DiarizationWhisperNLPFastAPI
10 Speech & Audio

Hinglish ASR

Speech recognition tuned for code-switched Hindi-English speech — the way India actually talks. Built for the low-resource, high real-world demand setting.

ASRHinglishLow-Resource
🛠️

// PROJECT CATEGORY 04

Full-Stack & Applied

Products people actually use — privacy-first apps, community platforms and dev tools.

5 projects

// MORE BUILDS

Full-Stack & Applied

📈 Vriddhi

Self-improving AI trading agent — a Hedge meta-allocator over 7 decorrelated strategies, walk-forward ML with a validation gate, and an adaptive risk overlay. Paper-trading research artifact.

PythonGBMWalk-ForwardRisk
Live Demo ↗ GitHub ↗

👻 Persona

Privacy-first social discovery & real-time chat — no accounts: ephemeral JWT sessions, E2E-encrypted self-destructing rooms, on-device location coarsening, QR/invite pairing.

ReactNode.jsWebSockets
Live Demo ↗ GitHub ↗

🔎 Lost & Found Network

Privacy-first community platform for lost & found items — smart matching engine, map view, anonymous chat.

Next.jsPostgreSQLLeaflet
Live Demo ↗ GitHub ↗

🧠 MindVault

Local-first private AI knowledge assistant — semantic search & chat with your docs. No telemetry; cloud-LLM option for demos.

PythonLocal LLMsRAG
Live Demo ↗ GitHub ↗

📡 LAN File Share

Zero-config LAN file sharing — upload, download & delete via web UI, with a QR code page for phone access. Single-file FastAPI.

FastAPIPythonNetworking
Live Demo ↗ GitHub ↗

// MY JOURNEY

Education & Background

2023 — Present

BSc (Hons.) Data Science & Artificial Intelligence

🎓 Indian Institute of Technology, Guwahati

Honours at one of India's most prestigious technical institutions. Core curriculum spans machine learning, deep learning, statistical inference, data engineering, algorithm design and AI systems — with research-backed projects and competitive coding along the way.

// LET'S CONNECT

Let's Build Something
Amazing

Looking for a data science intern, an AI collaborator, or just want to talk tech? I'm always open to great conversations and meaningful builds.