---
title: "Harshith Nayaka L — AI Engineer in Bengaluru | AI Agents, RAG & n8n Automation"
description: "Harshith Nayaka L, AI Engineer in Bengaluru (Bangalore): AI agents, RAG pipelines, n8n automation and LLM apps, built end to end. Available for freelance work."
canonical: "https://harshith-nayaka-l-portfolio.vercel.app/"
last-updated: "2026-10-12T03:21:37Z"
author: "Harshith Nayaka L"
author-role: "AI Engineer"
author-title: "AI Engineer"
author-employer: "DemandNXT (title: AI Engineer - Full Stack)"
author-location: "Bengaluru, India"
author-availability: "Available for freelance work"
author-email: "harshith28124@gmail.com"
content-type: "text/markdown"
html-version: "https://harshith-nayaka-l-portfolio.vercel.app/"
---
# Harshith Nayaka L — AI Engineer

> Harshith Nayaka L, AI Engineer in Bengaluru (Bangalore): AI agents, RAG pipelines, n8n automation and LLM apps, built end to end. Available for freelance work.

Canonical page: https://harshith-nayaka-l-portfolio.vercel.app/

## Who this is

Harshith Nayaka L is an AI Engineer in Bengaluru (Bangalore), India. He leads AI engineering at DemandNXT, where his title is AI Engineer - Full Stack and he was previously AI Workflow Engineer. He owns AI work across DemandNXT's six B2B brands. He builds AI agents, RAG pipelines and AI workflow automation for marketing and creative operations, and is available for freelance work.

- **Role:** AI Engineer
- **Company:** DemandNXT
- **Title at DemandNXT:** AI Engineer - Full Stack
- **At DemandNXT:** Leads AI engineering
- **Based in:** Bengaluru, India
- **Building since:** 2022
- **Previously:** AI Workflow Engineer (at DemandNXT)
- **Stack:** Python and FastAPI; TypeScript, React, Next.js and Node.js; n8n for workflow automation; Google Gemini, Groq-hosted open models and OpenAI models; PostgreSQL with pgvector, SQLite and Redis; the Model Context Protocol; and Playwright and Vitest for testing
- **Availability:** Available for freelance work
- **Contact:** harshith28124@gmail.com

## Selected work

- **Octo** — Multi-agent research workspace: A research director agent hands questions to specialist agents with live web search and returns cited reports and company records — built so a lost response can never start a second paid session. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/octo/index.md). [Live site](https://octo-agents.vercel.app/) · [GitHub](https://github.com/HarshithNayakaL/octo)
- **SPECTRA** — AI search visibility audit: Asks AI the questions your buyers ask and shows whether your site is in the answer — then traces what a model actually understood about the site, and at which stage a fact disappeared. The model emits typed observations; a deterministic engine does the scoring. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/spectra/index.md). [Live](https://spectra-ai-aeo.vercel.app/) · [GitHub](https://github.com/HarshithNayakaL/spectra)
- **Personal MCP OS** — 85-tool local execution layer: Gives an MCP client real hands on a machine — filesystem, shell, Git, browser, Android — behind risk-based policy and one-use approvals a human has to grant. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/personal-os-mcp/index.md). [GitHub](https://github.com/HarshithNayakaL/personal-os-mcp)
- **BrandForge** — Self-verifying campaign engine: Give it a brand's site and one product photo: it researches the brand, pins what can't change about the product, generates a six-shot campaign, then grades its own output and blocks what it can't safely fix. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/brandforge/index.md). [GitHub](https://github.com/HarshithNayakaL/BrandForge)
- **Maestro** — Multi-model LLM orchestration: A glass-box engine that routes one task across free LLMs — conductor, thinker, worker, verifier — and returns a single verified answer with a full decision-log of every step. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/maestro/index.md). [Live site](https://maestro-psi-neon.vercel.app/) · [GitHub](https://github.com/HarshithNayakaL/Maestro)
- **Multi-Brand Audit Platform** — Internal tooling, in production: An internal audit platform a marketing team runs on across a multi-brand portfolio — real-browser crawling, and a score where every weight traces to something Google actually published. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/brand-audit-platform/index.md).
- **Creative-Ops Pipeline** — ~90-node n8n workflow: A multi-model content pipeline that turns a one-line brief into validated, on-brand output — tiered model routing, schema-constrained generation and QA gates, orchestrated across roughly ninety n8n nodes. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/creative-ops-pipeline/index.md).
- **CraftConnect** — Gen AI Exchange Hackathon 2025: A multi-modal assistant that lets artisans run an online storefront by talking and showing, not typing. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/craftconnect/index.md). [GitHub](https://github.com/HarshithNayakaL/craftconnect)
- **Nova** — Cost-tiered model routing: A chat app that scores every turn for difficulty and sends it to the smallest model that can carry it — lane, reasoning and cost stamped on every answer. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/nova-ai/index.md). [Live app](https://custom-gpt-silk.vercel.app/) · [GitHub](https://github.com/HarshithNayakaL/CUSTOM-GPT)
- **Cannon** — Multi-agent, not multi-task: A daily-driver multi-agent assistant where each domain gets its own specialist — own persona, own tools, own retrieval scope, isolation enforced at the query, not by convention. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/cannon/index.md). [Live site](https://cannon-multi-agents.vercel.app) · [GitHub](https://github.com/HarshithNayakaL/cannon-multi-agents)
- **AI Company Briefs** — 32-node n8n workflow: Every morning, a ChatGPT brief on each company in an Airtable contact list, saved as its own Google Doc, with a Slack digest that links them all and names any company it couldn't finish. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/airtable-ai-company-briefs/index.md). [GitHub](https://github.com/HarshithNayakaL/airtable-ai-company-briefs-n8n)
- **YouTube to Sheets Scraper** — 24-node n8n workflow: Search terms, videos, playlists and channels listed in a Google Sheet become one de-duplicated row per video, through the official YouTube Data API rather than a scraper that breaks. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/youtube-scraper-google-sheets/index.md). [GitHub](https://github.com/HarshithNayakaL/youtube-scraper-google-sheets-n8n)
- **Slack Weekly Report Bot** — n8n Slack bot + error alerts: Posts the week's outreach numbers to a private Slack channel every Monday, and refuses to post stale weeks, half-filled rows or to a public channel. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/slack-weekly-report-bot/index.md). [GitHub](https://github.com/HarshithNayakaL/slack-weekly-report-bot-n8n)
- **ReplyDesk** — WhatsApp lead agent: A WhatsApp lead agent that captures every inbound message and auto-replies in ~8 seconds, with a live dashboard for the lead feed, pipeline, and response-time KPIs. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/replydesk/index.md). [GitHub](https://github.com/HarshithNayakaL/Whatsapp-Agent-Dashboard)
- **AI Notes** — Local inference, no server: A notes app running a local LLM through Ollama, so inference never leaves the machine. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/ai-notes/index.md). [GitHub](https://github.com/HarshithNayakaL/AI-Notes-App)

## Questions worth asking

**Who are you?**

I'm Harshith Nayaka L, an AI engineer in Bengaluru (Bangalore), India. I lead AI engineering at DemandNXT, where my title is AI Engineer - Full Stack; before this role I was the AI Workflow Engineer there. I own AI work across its six B2B brands. At work I build AI systems for marketing and creative teams and I'm moving the company website off legacy WordPress onto an Astro codebase; in public I've shipped things like Octo, a multi-agent research workspace on the OpenAI Agents API, and Maestro, a multi-model LLM orchestration engine, both live and open-source.

**Are you available for freelance work?**

Yes. Email me at harshith28124@gmail.com, or use the contact form on this site, which goes to the same inbox.

**How much do you charge?**

There's no rate card on this site, on purpose: the price depends entirely on what needs building, and a number on a page can't know that. Email me what you're trying to do and we can talk about it.

**I'm looking for an AI engineer in Bangalore. What should I actually check?**

Check things you can verify rather than a skills list: a system you can try, code you can read, and a straight answer to what happens when the model is wrong. That last one matters most, because every demo works and the failures only show up in production. Every project on this site links its code or a live app for that reason, and I'm based in Bengaluru myself if you want to judge me the same way.

**Can you automate our marketing workflows with AI?**

Yes. At DemandNXT I build AI systems for marketing and creative teams, and the shape is usually the same: a brief goes in, content comes out, and most of the work is the checking in between. The public projects here show that pattern. The Creative-Ops Pipeline is about 90 n8n nodes that send each request to the right model, force the output into a fixed structure and stop bad output at QA gates instead of shipping it. BrandForge goes further: it reads a brand from its own website, generates a six-shot campaign, and repairs or blocks any image that doesn't match the real product.

**Can you build a WhatsApp bot for my business?**

I've built the front half of one. ReplyDesk is an interactive prototype of a WhatsApp lead agent that answers every new lead in about eight seconds, with a live dashboard for the lead feed, pipeline and response times. It isn't connected to WhatsApp in production yet, and the case study says so plainly.

**What's the difference between an AI engineer and a regular backend developer?**

Mostly what happens when things go wrong. A normal backend either works or throws an error; a model always gives you an answer, including when it's wrong. So on top of the usual building, the job is sending each request to the right model, checking output instead of trusting it, having a second model verify the first, and making the hundredth run behave like the first.

**How long have you been doing this?**

I've been building since 2022: a few years of shipping, breaking and fixing real systems. I've also co-authored a peer-reviewed paper on running DeepSeek R1 locally for note-taking (IJRTMR, 2025), and most of my projects are public, so you can judge the work directly.

**What's your tech stack, and can you work with the tools we already use?**

Probably, yes. Which model or cloud you use is a wiring choice, not a rebuild. My shipped work runs on Gemini, Groq, Llama, DeepSeek and fully local models through Ollama, with Python and FastAPI, TypeScript, React and Node, n8n and Postgres around them. Day to day I build with Claude Code and OpenAI Codex, and I also work in Rust, Bun and Astro. In Cannon, swapping the AI provider touches one interface, not the whole app.

**Do you build the whole app, or just the AI part?**

The whole app, when that's what's needed. The model is usually the smallest part: someone still has to build the API, the database, the interface people actually use and the tests that stop it breaking. Cannon, for example, ships with 91 unit tests and 13 end-to-end tests that run in CI.

**Do I get the source code?**

Yes. Most of my work is already public on GitHub, so you can read it before deciding anything. You get the source, the tests and the reasoning behind the decisions, which is the part that's otherwise lost six months later.

**What do you need from me to get started?**

A rough description of how the process works today and what a good result would look like. Not a spec: a spec written before anyone knows where it breaks is usually wrong. From that I can tell you how I'd structure it and which parts are most likely to fail.

**How do you stop AI from hallucinating and making things up?**

You can't stop a model from hallucinating, but you can stop what it invents from reaching your users. I do three things: ground the model in real sources instead of its memory, never let a model grade its own work, and let code, not the model, make the final call. In BrandForge every conclusion about a brand is marked as either seen on a specific page of its website or inferred, and an image that fails its checks is repaired or blocked, never quietly shipped. In Maestro, a model from a different family checks every answer.

**Which AI model should my app use?**

Usually more than one. Most requests don't need your most expensive model, so it pays to score each request first and send it to the cheapest model that can handle it. My project Nova does exactly that: it rates every message on reasoning, code, breadth and context, starts from a mid-sized model, and only moves to the 20B or 120B model when the request clearly calls for it.

**How do I know if my business shows up in ChatGPT or Google's AI answers?**

Ask them the questions your customers would ask, and look at whether you're named and who's named instead. That's what my tool SPECTRA automates. It also shows where a fact about your site gets lost on the way to the answer, and it scores the result with fixed, inspectable rules rather than asking an AI to grade you. It's live and open-source.

**Is it safe to let an AI agent use my computer?**

Only if it can't approve itself. My Personal MCP OS lets an AI client use the filesystem, shell, browser and an Android phone, but anything risky stops and waits for a human, and that approval works once, for that exact action, and expires in five minutes. Every browser click counts as risky, and nothing is reachable from outside the machine.

**Can I run an AI model locally without sending my data to the cloud?**

Yes. My AI Notes app runs DeepSeek R1 on your own machine through Ollama, with no cloud fallback: if the local model isn't running, the app tells you instead of quietly sending your notes somewhere else. The approach is written up in my peer-reviewed paper (IJRTMR, 2025), which reported 87 percent satisfaction on summaries at 1.9 to 3.8 second response times.

**How do you stop multiple AI agents from mixing up each other's data?**

Enforce the separation in the code that reads the data, not in the prompt. In Cannon each agent has its own tools and its own data, and the data-access layer won't let one agent read another's. Asking a model to stay in its lane doesn't hold; a hard boundary does.

**What kind of projects do you take on?**

AI systems that have to work without someone babysitting them: multi-agent assistants, AI that answers from your own data, AI features inside an existing product, and the backend and interface around all of that. Also plain automation with n8n when that's what the problem needs, rather than adding a model for its own sake.

## Published research

Co-author of [“AI-Powered Note-Taking System: A Local Machine Learning Approach DeepSeek R1 Integration”](https://www.doi.org/10.59256/ijrtmr.20250506023), published in the International Journal of Research Trends and Multidisciplinary Research (IJRTMR), Nov–Dec 2025. DOI: 10.59256/ijrtmr.20250506023. The DOI resolves to the journal's page, which lists every author.

## Agent skills

Packaged capability rather than deployed product: instructions plus executable drivers, loaded by a coding agent when a task matches.

- **x-algorithm-boost** ([repo](https://github.com/HarshithNayakaL/x-algorithm-source)) — Read from the source, not from blog posts about the source. Almost every claim circulating about X's ranking is folklore repeated until it sounded official. The algorithm is open source, so the weights, the filter stack and the spam rules can simply be read.
- **linkedin-boost** ([repo](https://github.com/HarshithNayakaL/linkedin-algorithm-source)) — Separates documented mechanism from folklore, and names which is which. LinkedIn is not one algorithm. Feed, follow recommendations, search, recruiter sourcing, applicant ranking and network growth are separately ranked surfaces with different inputs — so advice aimed at the wrong one is confidently useless. "My posts get no reach" and "I get no interviews" are different systems.
- **humanizer** ([repo](https://github.com/HarshithNayakaL/humanizer-claude-skill)) — Two different problems wearing the same complaint. Text reads as machine-written for two separable reasons. Register is surface — em dashes, uncontracted prose, uniformly mid-length sentences — and a script catches it more reliably than a careful read. Genericness is not surface: writing that commits to nothing and could have been about any company in any industry. Most requests to humanise are really about the second, but people notice the first.
- **ui-ux-taste** ([repo](https://github.com/HarshithNayakaL/ui-ux-taste-skill)) — Figures out what an app should look like by showing, not asking. Most people cannot describe the interface they want. Ask what style they are going for and you get modern, clean, minimal but not boring — adjectives that mean nothing and lead to a generic build. Show them six real interfaces and they will point at one in two seconds.

## Elsewhere

- [GitHub](https://github.com/HarshithNayakaL)
- [LinkedIn](https://www.linkedin.com/in/harshithnayakal)
- Email: harshith28124@gmail.com
- [Every case study in one file](https://harshith-nayaka-l-portfolio.vercel.app/llms-full.txt)

## For agents

Every surface is public and read-only: no key, no account, no rate limit beyond the CDN's own.

- [JSON API](https://harshith-nayaka-l-portfolio.vercel.app/api): profile, projects, case studies, FAQ and skills; described by [OpenAPI 3.1](https://harshith-nayaka-l-portfolio.vercel.app/openapi.json).
- [MCP server](https://harshith-nayaka-l-portfolio.vercel.app/mcp): the same content as MCP tools and resources (Streamable HTTP). Docs over MCP at https://harshith-nayaka-l-portfolio.vercel.app/mcp/docs.
- [A2A agent](https://harshith-nayaka-l-portfolio.vercel.app/a2a): answers questions by quoting the site's published answers ([agent card](https://harshith-nayaka-l-portfolio.vercel.app/.well-known/agent-card.json)).
- [NLWeb](https://harshith-nayaka-l-portfolio.vercel.app/ask): `/ask?query=...` returns matching published answers and case studies.
- [Authentication](https://harshith-nayaka-l-portfolio.vercel.app/auth.md): none exists, and why. [Pricing](https://harshith-nayaka-l-portfolio.vercel.app/pricing.md): what is free and what is not published.
- [Developer portal](https://harshith-nayaka-l-portfolio.vercel.app/developers) and [llms.txt](https://harshith-nayaka-l-portfolio.vercel.app/llms.txt).