--- title: "Harshith Nayaka L — AI Engineer, Full-Stack | Bengaluru" description: "Full-stack AI Engineer in Bengaluru (Bangalore), India. I build AI agents, RAG pipelines and LLM apps, plus the backends and interfaces around them." canonical: "https://harshith-nayaka-l-portfolio.vercel.app/" last-updated: "2026-08-22T04:29:48Z" author: "Harshith Nayaka L" content-type: "text/markdown" html-version: "https://harshith-nayaka-l-portfolio.vercel.app/" --- # Harshith Nayaka L — AI Engineer, Full-Stack > Full-stack AI Engineer in Bengaluru (Bangalore), India. I build AI agents, RAG pipelines and LLM apps, plus the backends and interfaces around them. Canonical page: https://harshith-nayaka-l-portfolio.vercel.app/ ## Selected work - **Creative-Ops Pipeline** — Flagship: A multi-model content pipeline that replaces hours of manual production with validated, on-brand output. [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) - **Maestro** — Multi-model 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) - **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) - **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) - **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) - **BlogSpace** — Live & deployed: A full-stack blogging platform with JWT auth and an admin panel, live in production. [Case study](https://harshith-nayaka-l-portfolio.vercel.app/work/blogspace/index.md). [GitHub](https://github.com/HarshithNayakaL/blogspace-internship) - **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 **How do you stop an LLM from marking its own homework?** You do not let the model that produced the answer be the model that judges it. Maestro routes every worker output through a verifier on a deliberately different model family, because a model asked to grade its own work carries a documented self-preference bias of roughly 10 to 25 percent. The verifier returns a pass or fail verdict with the issues it found, and a fail triggers one bounded retry rather than an unbounded loop. The whole decision-log is replayable, so you can see which model said what and why. **How do you decide which model should answer a request?** By scoring the request before spending on it. Nova reads every turn on four axes — reasoning, code, breadth and context — produces a deterministic complexity score, then lets a small arbiter model confirm or overrule that reading. The counterintuitive part is that the middle tier is the default, not the cheapest: a request has to earn its way down to the 20B model by being demonstrably trivial, or up to the 120B by being demonstrably hard. Routers that climb from cheapest upward park almost everything in the fast lane and answer it badly. **How do you keep multiple agents from leaking context into each other?** By enforcing isolation at the query rather than trusting convention. In Cannon each domain agent has its own persona, its own tools and its own retrieval scope, and no agent can read another's data — the boundary is in the data access layer, not in a prompt asking the model to behave. That is the opposite architecture to an orchestrator like Maestro, where roles are meant to collaborate on one task. It ships with 91 unit tests and 13 end-to-end specs that run in CI without secrets. **Can you run an LLM without sending data to a cloud provider?** Yes, and with no cloud fallback at all, which is the part that matters. AI Notes runs DeepSeek R1 locally through Ollama for summarisation, keyword extraction and writing assistance; when the local model is not running the app says so rather than quietly shipping your notes somewhere else. The architecture is documented in a peer-reviewed paper I co-authored (IJRTMR, 2025, DOI 10.59256/ijrtmr.20250506023), which reported 87 percent satisfaction on summarisation at 1.9 to 3.8 second response times. **What kind of work do you take on?** AI systems that have to run unattended: multi-agent assistants, retrieval pipelines grounded in real data, LLM features inside an existing product, and the backends and interfaces around them. Also the automation layer when that is what the problem actually needs — n8n and integration pipelines rather than a model for its own sake. **How do you keep an AI system reliable in production?** By assuming the model can be wrong at every step. Output is requested against a schema and then validated anyway, because a schema request is a hope rather than a guarantee. Failures are caught at the boundary they happen at, not three steps downstream. Maestro routes work through a verifier whose model family is deliberately different from the worker's, so the check is independent. Cannon falls back across providers behind one typed interface, so an outage degrades the answer instead of breaking the app. **Do you work with the models and infrastructure we already use?** Yes — provider choice is a wiring decision, not an architectural one. Shipped work runs on Gemini, Groq, Llama, DeepSeek and fully local inference through Ollama, on Google Cloud and Hugging Face, with FastAPI, Node, Postgres and React around it. Cannon in particular is built so swapping a provider touches one interface rather than the application. **Do you build the whole application, or only the AI part?** The whole thing, when that is useful. The model is usually the least of it: someone still has to build the API, the data layer, the interface a person actually uses, and the tests that stop it regressing. Cannon ships with 91 unit tests and 13 end-to-end specs that run in CI without secrets. **Do I get the code?** Yes. Most of the work above is public on GitHub and you can read it before deciding anything — that is deliberate. You get the source, the tests, and the reasoning for the decisions that would otherwise be invisible six months later. **What do you need to get started?** The rough steps of the process as it works today, and what a good outcome looks like. Not a spec — a spec written before anyone understands the failure modes is usually wrong. From that I can tell you what the structure should be, and where the parts that will actually break are. ## 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)