Bengaluru · 12.97°N 77.59°E · 2026

Rishab Dugar

Senior Data Scientist building agentic AI systems and production data platforms that run at scale — not in a demo.

I spend my days making language models behave like dependable software: retrieval that actually retrieves, agents that know when to stop, and pipelines that survive the night without paging anyone. Triple cloud certified, three years of production systems, and a long-standing habit of reading about cosmology when the deploy is running.

Senior Data Scientist · V4C AWS DVA-C02 GCP Pro ML Engineer Azure AI-102 Top 1% LinkedIn Programming Voice

Move the pointer — or drag a finger — the sky answers. Tap for a spark.

Observations

What the work moved.

~90% AI cost reduction per interaction, and the same again on QA effort
<500ms end-to-end latency on a live voice + chat AI agent
~200x cheaper analytical queries, with 10x columnar compression
45% retrieval precision uplift from hybrid search and re-ranking
96.36% accuracy on an LLM-generated test-case suite
60% MTTR reduction from centralised logging and automated runbooks

Trajectory

Where it shipped.

V4C

2026 — Present

Senior Data Scientist

Applied AI and data science on production systems — modelling, retrieval and agentic workflows that go in front of real users. Recently started; details here as they ship.

Accenture

Aug 2023 — 2026 · Bengaluru / Kolkata

Data & AI Senior Analyst

  • Built autonomous multi-agent systems (MCP, A2A, LangGraph, Strands SDK) with agentic RAG — hybrid search plus re-ranking — for enterprise clients.
  • Engineered a single-brain voice and chat AI agent unifying IVR, WebRTC and chat at sub-500 ms end-to-end latency.
  • Designed a serverless GenAI analytics pipeline on AWS — real-time and batch, Parquet on Athena — cutting per-interaction AI cost ~90%.
  • Owned observability and incident response for the platforms above, and ran GenAI workshops for 30+ engineering interns.

Technical writing

2022 — Present

GeeksforGeeks · Medium · dev.to · LinkedIn

4k+ monthly reads across platforms and 13k+ followers, mostly on the parts of applied AI that usually go undocumented. Top 1% LinkedIn Programming Voice, FY 2024–25.

Missions

Built outside the ticket queue.

Gyanshala — Jain education web apps for children

Gyanshala — Jain learning, as a game

A free, ad-free suite of learning apps teaching Terapanth history, the lineage of Acharyas and tirtha geography to children aged 4–13 — a tradition taught orally in classrooms, now reachable by any family anywhere at zero cost.

Every app is a single dependency-free HTML file: no build pipeline, no backend, no per-user cost. Bilingual Hindi/English quizzes and matching games, a global error-recovery boundary, muted-by-default Web Audio, and a WhatsApp-native sharing loop. A production freeze was traced with CDP heap and listener profiling across 1,000+ automated interaction cycles and fixed.

Solo-designed — architecture, UX and content — and iterated from real parent and child feedback into stable public releases. Live and in use.

Vanilla HTMLWeb AudioCloudflare Pages i18nZero deps
gyanshala.pages.dev →
GenAI Stock Analysis Agent

GenAI Stock Analysis Agent

A multi-agent research desk: cross-validates 1,000+ daily data points across five APIs with integrity checkpoints, holding >99.7% data accuracy, and turns an afternoon of report writing into under five minutes.

LangGraphLangChainGemini yFinanceAlpha Vantage
GitHub →
AI invoice and commerce platform

AI Invoice & Commerce Platform

Full-stack commerce system with LLM structured extraction, voice-to-invoice speech-to-text and ECDSA digital signatures. Thread-safe in-memory caching cut disk I/O ~95%; it runs on a self-hosted Linux homelab in a hybrid cloud setup.

FastAPIReactGeminiJWTECDSA
Live site →
Multi-agentic LLM recommendation orchestrator

Multi-Agent Recommendation Orchestrator

Several specialised agents behind one question — Reddit and web retrieval, advanced retrievers and RAG, arbitrated into a single recommendation instead of five competing ones.

LangGraphAzure OpenAIRAGPRAW SerperStreamlit
GitHub →
SLM fine-tuning pipeline

SLM Fine-Tuning Pipeline

Fine-tuned LLaMA 3 with 4-bit PEFT-LoRA on a free Colab tier, then published the whole workflow — evaluation metrics, benchmarks and the failure modes people normally leave out of the tutorial.

UnslothPEFT-LoRALLaMA 3Colab
Read on Medium →
Voice-to-voice GenAI multi-LLM bot

Voice-to-Voice Multi-LLM Bot

Real-time spoken dialogue with a switchable model behind it — Whisper in, any of several LLMs in the middle, speech back out. Built to compare backends on the same conversation rather than on a benchmark.

WhisperOpenAIGeminiLangChain Hugging Face
GitHub →
UAV for precision agriculture

UAV for Precision Agriculture

Simulated spraying drone with coverage path planning at 84% crop coverage precision, using U-Net and R2U-Net semantic segmentation to find land boundaries on its own.

PyTorchU-NetROSGazeboRaspberry Pi
Deck →
GemSign invoice generator

GemSign — Signed Invoices

PIN-based TOML authentication at sub-500 ms, and cloud PPTX→PDF conversion of client decks in about three seconds, synced straight back to Drive.

StreamlitGoogle Drive APIPython
Live app →
Green Leaf Gourmet restaurant POS

Green Leaf Gourmet

A full restaurant point-of-sale, shipped the awkward way on purpose: a single Windows EXE with a bundled JRE, so a restaurant with no IT desk can install it by double-clicking.

JavaSwingMySQLiTextInno Setup
GitHub →
ChronoShell, the 1888 edition of this site

ChronoShell — this site, 1888 edition

A Victorian laboratory you can actually use: a working command shell inside the machine, 37 measured hotspots on a looping video plate, a journal that turns its own pages, and an assistant running tf-idf retrieval with a deterministic FAQ gate on Cloudflare Workers — no vector database anywhere.

Vanilla JSCSS 3DCloudflare Pages WorkersRAG
Take the tunnel →

Smaller things that did not need a card: a bulk subtitle scraper written twice, once with libraries and once without, to show the difference; and an over-engineered birthday animation in pure CSS, which is exactly as serious as it sounds. Everything else lives on Linktree.

Elsewhere

The rest of the trail.

Instruments

What I reach for.

Languages

PythonSQLJavaJavaScriptC++

GenAI & Agents

Multi-agent systemsMCPA2ALangGraph LangChainStrands SDKGoogle ADKLlamaIndex Agentic RAGHybrid searchRe-ranking

ML & Deep Learning

PyTorchTensorFlowPEFT-LoRAUnsloth 4-bit quantizationHugging FaceTransformers Semantic segmentation

Cloud & AI services

AWS BedrockNova SonicLambdaAgentCore KinesisGlueAthenaQuickSight GCP Vertex AIAzure OpenAIDockerCI/CD

Data & storage

DynamoDBPineconeChromaDBOpenSearch RedisParquetText-to-SQLKafka

Reliability

Distributed tracingCloudWatchKibana Circuit breakersRetriesIncident response

Credentials

Verified, not claimed.

AWS Certified Developer Associate badge AWS Certified Developer Associate · DVA-C02 Verify
Google Cloud Professional Machine Learning Engineer badge Professional ML Engineer Google Cloud Verify
Microsoft Azure AI Engineer Associate badge Azure AI Engineer Associate · AI-102 Verify
LeetCode Knight badge LeetCode Knight 1983 · top 2.84% · 450+ solved Verify

Also: B.Tech, Electrical Engineering, NIT Durgapur (CGPA 8.72) · Amazon ML Summer School 2022, top 0.5% of 19,000+ · TCS CodeVita S10, global rank 1,979 of 136,000+ · CodeChef 4-star, max 1892 · NPTEL C Programming (IIT KGP), 100% · AIR 1, Jain Vidya Exams I, II and V.

Lockwood

Ask my assistant anything.

Lockwood here — Rishab's assistant, and firmly on his side. Ask about his work, his stack, or what he actually shipped. I answer from his résumé, so I will not make things up.

Signal

Let's build something that stays up.

Open to applied-AI, data science and platform work. Fastest reply is email; LinkedIn works too.