Nishi Kant Chandra

AI Engineer · Data Analytics · Generative AI · Full-Stack AI Developer

Based between New Delhi, India and Moscow, Russia

I build AI systems that run in production and teach the skills needed to build them. That means retrieval-augmented assistants, agents and analytics pipelines for clients, and live cohort courses for people moving into AI work. Most of what I know came from shipping and maintaining this site: 231 free tools, nine languages, and a build that refuses to publish a claim the code cannot honour.

Hire me for a projectLearn from meWhatsApp
[email protected]GitHub

HSE University, Moscow

M.Sc. Business Analytics and Big Data Systems

O.P. Jindal Global University

MBA Business Analytics — O grade

Google

Data Analytics Professional Certificate

Vanderbilt University

Advanced Prompt Engineering

Three things, one person

What I do now

The teaching, the client work and the tools are the same skill pointed in three directions. Each one keeps the other two honest.

Teach

Five live cohort courses, from AI for marketers through to RAG, agents and production LLM engineering. Small batches, real projects, no placement promises.

See the courses →

Build

Chatbots, agents, RAG systems, analytics dashboards, technical SEO and WordPress recovery. Fixed scope, written figure before anything starts, source handed over.

See the services →

Ship

231 free browser tools, maintained across nine languages with a build pipeline that refuses to publish a claim the code cannot honour.

Use the tools →

Selected work

Five projects, written up properly

Problem, approach, the decisions that mattered, and what it produced. Four have live demos you can open right now; the fifth is the site you are reading.

Yoga Pose Correction Assistant

Real-time computer vision that never sends your camera anywhere

Live demo
The problem
Correcting posture needs someone watching you, and the obvious way to build that — stream video to a server for inference — asks people to upload footage of themselves exercising at home. Most will decline, and they are right to.
Approach
Pose estimation runs entirely in the browser. MoveNet Thunder through TensorFlow.js reads seventeen keypoints per frame, joint angles are compared against reference poses, and corrections are spoken through the Web Speech API while the skeleton overlays live video.
Decisions that mattered
  • Client-side inference over a server API — the video never leaves the device, so there is no upload to consent to and no footage to retain
  • Thunder over Lightning: the accuracy difference matters for joint angles, and the frame budget still holds on a mid-range laptop
  • Angle thresholds per pose rather than a trained classifier — five poses did not justify a model, and thresholds can be explained to the user
Result
Five poses detected and corrected in real time, with visual and voice feedback, running on an ordinary laptop webcam. This was the practical half of the MBA thesis.
TensorFlow.jsMoveNet ThunderReactTypeScriptOpenCVWeb Speech API

Motorcycle Match

A recommender that tells you why

Live demo
The problem
Recommendation engines that cannot explain themselves are not trusted on considered purchases. Nobody spends two lakh rupees because a black box suggested it.
Approach
A rule-based filter narrows the field on hard constraints (budget, licence, use case), content-based similarity ranks what survives, and a rider-persona classifier frames the result. Every recommendation carries the reasons that produced it.
Decisions that mattered
  • Explainability designed in from the start, not bolted on — the ranking features are the explanation, so the two cannot drift apart
  • Hard constraints filter before scoring, so nothing unaffordable or unlicensable can be ranked highly and then apologised for
  • Personas as a presentation layer over the scores, not as the scoring itself, so a rider who does not fit a persona still gets sensible results
Result
A working recommender where each suggestion states the constraints it satisfied and the attributes it matched on.
PythonContent-based filteringRule enginesExplainable AI

Air Quality Forecaster

Deep learning on a problem Delhi feels personally

Live demo
The problem
Published AQI figures tell you about the air you already breathed. The useful question is what tomorrow looks like, and that is a sequence problem, not a lookup.
Approach
A GRU network trained on historical AQI series produces short-horizon forecasts, served through an interactive dashboard that accepts a CSV and charts observed against predicted.
Decisions that mattered
  • GRU over LSTM — fewer parameters, comparable accuracy on this series length, and it trains in a fraction of the time on the hardware available
  • CSV in, chart out: the dashboard takes the data a user actually has rather than assuming a live sensor feed
  • Forecast horizon kept short and stated plainly, because a confident seven-day AQI number would be fiction
Result
A deployed forecaster with an interactive dashboard, visualising predicted against observed values over the input series.
PythonGRU / RNNTime-series forecastingData visualisation

Polingo

Speak, get corrected, sound native — real-time voice AI that had to be cheap enough to give away

Live demo
The problem
Speaking is the part of a language nobody practises, because practising it needs a patient partner who corrects you. AI can be that partner, but the obvious build — stream audio to a paid speech API, send the transcript to a frontier model, synthesise a reply — costs real money on every single turn of every single conversation. At that unit cost the product only works for people who can already afford a tutor, which is the wrong half of the market.
Approach
A full-stack platform where the cost model is a design constraint rather than an afterthought. Voice runs over WebSockets as streaming audio through speech-to-text and text-to-speech, with more than one provider behind each so the cheapest capable tier takes the call; the coaching layer routes between models by how hard the turn actually is and caches semantically, so near-identical corrections are not paid for twice. Around that sits an A1–C2 curriculum with quizzes and spaced-repetition vocabulary, roleplay scenarios, pronunciation scoring, peer-to-peer conversation rooms over WebRTC, and subscription billing.
Decisions that mattered
  • Multi-provider speech and language routing rather than one vendor — the free and low tiers of several providers cover the ordinary turn, and no single provider outage or price change can take the product down or make it unaffordable
  • Semantic caching on the coaching layer, not just exact-match — learners make the same mistakes as each other, so the second person to mispronounce a word should not cost what the first did
  • Peer-to-peer conversation goes direct over WebRTC with only a signalling server in the middle, so human-to-human practice adds bandwidth cost rather than per-minute AI cost
  • Spoken correction over written correction: the feedback is heard, because a learner reading a correction is no longer practising the skill being corrected
Result
A live platform at polingo.app: real-time spoken conversation and correction across 38 languages, focused on English ↔ Russian, with a structured curriculum, progress tracking and paid plans behind it. More detailed technical information and the system architecture are not shared publicly, for security reasons.
Next.js 16React 19TypeScriptTailwind CSSFastAPIPostgreSQLRedisWebSocketsWebRTCDocker

NKable.com — this site

A build pipeline that refuses to publish a dishonest sentence

The problem
A tools site with hundreds of pages accumulates copy nobody re-reads. An audit of this one found tool pages promising exports, uploads and saved history that did not exist, twelve tools filling their results in with random data, four SEO configs describing a different tool entirely, and a contact form that logged every message to a console call the production build strips. Every enquiry ever submitted was discarded while the sender was told it had arrived.
Approach
Fix the pages, retire what could not be fixed honestly, then make the failure mode impossible to repeat. Three build guards now run before every deploy: one greps each page for the capability its copy claims, one holds nine locale bundles to exact key and placeholder parity, one proves every live tool is reachable from the site chrome and that no chrome link is dead.
Decisions that mattered
  • Retire rather than fake: fourteen tools that could not work in a browser — DNS lookups, port scans, TLS inspection — were withdrawn instead of continuing to invent output. They still return 200 so inbound links survive, and are excluded from the sitemap and every agent manifest
  • Published commitments are asserted against the code: the 90-day retention period on /privacy is compared at build time against the value the backend sweep actually enforces, and the build fails if they disagree
  • Guards are tested by planting the violation first, then confirming the clean run — an untested guard is a guard that passes vacuously
Result
231 tools live across ~2,900 prerendered pages in nine languages, an agent-native surface (MCP endpoint, llms.txt, RFC 9421 signatures, .well-known manifests), a daily retention sweep, and full sitemap coverage. The audit and every fix are documented in the repo.
Next.jsTypeScriptFastAPIPostgreSQLDockerCloudflareGitHub Actions

Skills

Banded by how well I actually know them

A flat list of forty keywords tells you nothing, because everyone's list looks the same. This one is ordered by depth, including the band that admits to being shallow.

Build with daily

Used on most working days, on this site and on client work.

PythonSQLTypeScriptReactNext.jsTailwind CSSFastAPIGit

Shipped to production

Built, deployed and maintained for real users.

RAG systemsPrompt engineeringAI agentsPower BIDockerREST API designTechnical SEO

Applied in projects

Used end to end on the projects below, not in a long-running production system.

TensorFlowTensorFlow.jsComputer visionRecommender systemsGRU / time-series forecastingOpenCV

Working knowledge

Enough to build with and to know where the edges are.

Vector databasesCloudflare WorkersCeleryPostgreSQLGitHub Actions

Languages

Hindi — Native

English — C1

Russian — A1

Experience

May 2022 — October 2024

Product Support Specialist II

Experience.com

Tier-2 technical support for a SaaS platform, working directly with enterprise clients and alongside the engineering and product teams.

  • Supported 100+ SaaS clients at tier 2, holding 95%+ satisfaction
  • Ran client onboarding and training sessions
  • Debugged API, HTML and CSS issues against a live platform
  • Wrote the SQL and Excel reporting the support team ran on
  • Reduced inbound ticket volume by roughly 25–30% through documentation and fixes
  • Mentored new joiners and fed recurring problems back into the product roadmap
SQLExcelJiraREST APIsHTMLCSS

January 2015 — May 2022

Independent Retail & E-Commerce Operations

Amazon Marketplace (self-employed)

Ran a books and software retail business end to end — storefront, listings, advertising, inventory, pricing and customer support.

  • Built and operated the storefront, listings and advertising campaigns
  • Grew revenue by roughly 30% through listing and pricing optimisation
  • Maintained 100% order fulfilment across the period
  • Managed inventory and sales analytics without a team
Amazon Seller CentralAmazon AdsExcelSales analytics

Education

2024 — 2026

M.Sc. Business Analytics and Big Data Systems

Higher School of Economics (HSE University)

Moscow, Russia

Completed July 2026, studied in Moscow while teaching cohorts across India.

Machine Learning · Neural Networks · Deep Learning · Predictive Analytics · Recommender Systems · Data Engineering · Big Data Architecture · Applied Data Science · Decision Support Systems

2024 — 2026

MBA — Business Analytics

O.P. Jindal Global University

New Delhi, India

Final grade 7.375 / 8.0 (O). Thesis: Computer Vision-Based Pose Estimation for Quantitative Assessment of Postural Accuracy in Yoga Practice — the research behind the pose-correction assistant below.

Certifications

Google Data Analytics Professional Certificate

Google · 2025

Advanced Prompt Engineering

Vanderbilt University · 2025

English Upper Intermediate (B2.1)

Università di Napoli Federico II · 2025

How I work

Four principles, each with something on this site backing it

Anyone can write these as adjectives. Every one below points at a file in the repository that enforces it.

01

If it can be checked, it is checked in CI

Five scripts run before every build: copy against capability, locales against each other, links against routes, marketing claims against the catalogue, and published personal data against a withheld list. Reviews miss things. Scripts do not get tired.

02

An honest no beats a promise nobody implemented

"There is no export here — copy it instead" is better copy than an export button that was never built. The build guard on this site enforces exactly that, and it has failed builds over it.

03

Data you do not keep cannot leak

Usage rows are deleted after 90 days by a daily sweep, and the number on the privacy page is compared against the number in the backend config at build time, so the site cannot publish a commitment the code does not keep.

04

Ship it, then measure it

231 tools live, 231 in the sitemap, nine locales at exact key parity. Every one of those numbers is asserted by a script rather than remembered.

The source for all of it, including the audit that found fourteen tools returning fabricated output and retired them, is on GitHub.

About me, briefly

What does Nishi Kant Chandra build?

AI systems that run in production: retrieval-augmented assistants over private documents, agents with bounded tool use, real-time voice applications, analytics pipelines and dashboards, and full-stack web applications. Four have live demos — Polingo, a real-time voice language-learning platform, plus pose estimation, a recommender and an AQI forecaster.

Where is he based?

Between New Delhi, India and Moscow, Russia. Client work and teaching are both delivered remotely, on Indian Standard Time.

What is his background?

An MBA in Business Analytics from O.P. Jindal Global University with an O grade, an M.Sc. in Business Analytics and Big Data Systems from HSE University Moscow completed in July 2026, Google's Data Analytics certificate, and two and a half years of tier-2 SaaS product support for enterprise clients.

Is NKable.com his own work?

Yes — the platform, its 231 tools, the FastAPI backend, the nine-locale setup and the build guards that keep its copy honest were all built and are maintained by him. It is written up as the last case study on this page.

Is he available for projects?

Yes, for fixed-scope engagements in AI engineering, analytics, technical SEO and web development. Availability and price are agreed on a short call before anything starts.

Does he teach?

Yes. Five live online cohort courses run from beginner AI marketing through to AI engineering with RAG, agents and production deployment.

Want to work together, or learn this?

A short call is the fastest way to find out whether I am the right person for it. If I am not, I will say so — that is a cheaper conversation for both of us than a project that should not have started.

Book a callWhatsAppEmail