# Deep Parmar — AI Engineer > Portfolio and blog of Deep Parmar — AI Engineer, CTO & co-founder of Sunbots Innovations, and founder of Xwits. Based in Ahmedabad, India. Deep specialises in prompt engineering, context engineering, harness engineering for LLMs, computer vision, edge AI, browser-native AI, and building production AI products. He is the creator of Dhiya NPM (client-side RAG, published on npm as `dhiya-npm`) and SmartON — AI-powered smart glasses used by 17,000+ blind and visually-impaired users. ## Expertise - **Prompt Engineering**: Designing LLM inputs for production — zero-shot, few-shot, chain-of-thought, system prompts, ReAct pattern, output structuring, prompt versioning and evaluation - **Context Engineering**: Designing what goes into LLM context windows — RAG context design, context compression, conversation memory strategies, context poisoning prevention, context budget management - **Harness Engineering**: Building production infrastructure for LLMs — routers, input/output guardrails, evaluation harnesses, agent orchestration, observability, retry and fallback logic - **AI Agents & Agentic AI**: AI agent architecture, agentic AI design philosophy, autonomy spectrum, multi-step reasoning, tool use, agent safety, supervised vs autonomous agents - **Computer Vision**: Object detection, YOLO, TFLite on Android, real-time edge inference, OCR, scene understanding - **Edge AI**: Deploying AI on NVIDIA Jetson Nano, WebGPU inference, client-side ML, on-device inference - **Browser AI / Client-Side RAG**: Running LLMs in the browser with Dhiya NPM, Transformers.js, IndexedDB vector stores, WebGPU, privacy-first AI - **Multilingual AI for India**: Building AI for Indian languages (Hindi, Gujarati, English), multilingual routing, code-switching, MIRA assistant - **Gen-AI Production Strategy**: RAG vs fine-tuning decisions, production AI stack, why AI prototypes fail, gen-AI deployment choices - **AI Consulting & Product Building**: Hiring AI engineers, AI project scoping, custom AI vs API cost analysis, building AI products in 90 days - **AI in India / Career**: AI CTO role, building AI startups in Tier-2 India, learning AI while running a company, open source as career strategy ## Pages - [Home](https://deepap.dev/): About Deep Parmar, featured projects (SmartON, Marketing Autopilot, XwFin, skibdrip), contact form. - [About](https://deepap.dev/about): Long-form biography of Deep Parmar — AI engineer and CTO in Ahmedabad, India. - [Projects](https://deepap.dev/projects): AI projects by Deep Parmar — SmartON, Marketing Autopilot, XwFin, skibdrip, Dhiya NPM. - [Contact](https://deepap.dev/contact): Contact Deep Parmar — email, phone, and social profiles. - [Blogs](https://deepap.dev/blogs): All 78 articles on AI engineering, production ML, browser AI, career, edge AI, and AI in India. ## Blog Posts — AI, Without the Hype (Plain-English Series) - [The Next 5 Years of AI — Without the Sci-Fi](https://deepap.dev/blogs/next-5-years-ai-no-scifi): Expect AI to get more useful, more boring, and more invisible — woven into ordinary tools — not a sudden sci-fi takeover by conscious machines. - [How a Small Business Can Actually Use AI (Without Wasting Money)](https://deepap.dev/blogs/ai-for-small-business-india): Start with one painful, repetitive task. Use an off-the-shelf tool before building anything. Measure whether it saved time or money before expanding. - [AI That Speaks Indian — Hindi, Gujarati, and the Rest](https://deepap.dev/blogs/ai-that-speaks-indian-languages): AI now works in Indian languages — imperfectly but usefully. That, more than English chatbots, is what will bring AI to the next half-billion Indians. - ["Will AI Take My Job?" — An Honest Answer for India](https://deepap.dev/blogs/will-ai-take-my-job-india): In most cases AI automates tasks, not whole jobs. The person who uses AI well tends to replace the one who doesn't — rather than AI replacing both. - [AI That Runs on Your Phone, Not in Someone's Cloud](https://deepap.dev/blogs/ai-on-your-phone-on-device): On-device AI runs the model locally on your phone, laptop, or browser — so it works offline, keeps your data private, and costs nothing per use. - [Where Does Your Data Go When You Use AI?](https://deepap.dev/blogs/where-your-data-goes-ai): With most cloud AI tools, what you type or upload is sent to a company's servers and may be stored or used to improve their models. On-device AI keeps it on your machine. - [Talking to AI So It Actually Helps (Prompting for Normal People)](https://deepap.dev/blogs/prompting-for-normal-people): You get far better results from AI by telling it who it is, giving it context, showing one example, and stating the format you want. Five simple moves, no jargon. - [Why AI Confidently Lies — and How to Catch It](https://deepap.dev/blogs/why-ai-lies-hallucinations): AI "hallucinates" because it predicts plausible-sounding text, not verified truth. When a confident wrong answer is statistically likely, it produces it anyway. - [How AI Actually "Thinks" — It's Fancy Autocomplete, Not Magic](https://deepap.dev/blogs/how-ai-thinks-fancy-autocomplete): AI does not think. It predicts the next word, over and over, based on patterns it learned from huge amounts of text. There is no understanding, and no mind. - [What AI Actually Is in 2026 (Minus the Hype)](https://deepap.dev/blogs/what-ai-actually-is-2026): AI in 2026 is software that learns patterns from data and uses them to predict or generate things — minus the magic and the menace. Part 1 of a plain-English series on what AI really is, by someone who builds it. ## Blog Posts — Browser AI & Client-Side ML - [Browser AI in 2026: What's Next for Dhiya NPM and Client-Side Intelligence](https://deepap.dev/blogs/browser-ai-dhiya-npm-2026-roadmap): Eight months after launching Dhiya NPM, here is where browser AI actually stands, what surprised us, and what the 2026 roadmap looks like for client-side inference. - [Dhiya NPM — No-Cost AI for the Web: Build RAG Bots That Run Entirely in the Browser](https://deepap.dev/blogs/dhiya-npm-no-cost-ai-web): Introducing Dhiya NPM — a client-side RAG framework that runs entirely in your browser. Build ChatGPT-style assistants without servers, API keys, or recurring cost. - [Privacy-First AI: Why Client-Side Inference Matters](https://deepap.dev/blogs/privacy-first-client-side-ai): Every API call leaks data. On-device AI keeps user data on the device. Here's why this matters and how to build for it. - [IndexedDB as a Vector Store: Browser Embeddings That Persist](https://deepap.dev/blogs/indexeddb-vector-store-embeddings): IndexedDB can store and query embeddings entirely in the browser. Here's how Dhiya NPM implements a client-side vector store with no backend required. - [Transformers.js: Running LLMs in the Browser](https://deepap.dev/blogs/transformers-js-browser-llm): Transformers.js brings Hugging Face's model ecosystem to the browser. Here's how to use it for embeddings, classification, and small LLMs. - [WebGPU for AI Inference: A Web Developer's Guide](https://deepap.dev/blogs/webgpu-ai-inference-guide): WebGPU makes GPU-accelerated AI inference practical in the browser. Here's what web developers need to know to use it effectively. - [How to Build a RAG Chatbot with Dhiya NPM](https://deepap.dev/blogs/build-rag-chatbot-dhiya-npm): A complete tutorial: from npm install to a working RAG chatbot that runs entirely in your browser, with no server and no API key required. - [Client-Side RAG: Running AI in Your Browser](https://deepap.dev/blogs/client-side-rag-browser-ai): A RAG chatbot with no server, no API keys, and no monthly bill — running entirely in your browser. Here's how it works. ## Blog Posts — LLM & AI Engineering - [Prompt Engineering: A Practical Guide for 2026](https://deepap.dev/blogs/prompt-engineering-guide): Prompt engineering is the art of communicating intent to a language model. In 2026 — with newer models like Claude Fable 5 — it's shifting from controlling the model to collaborating with it. Here's the practical guide. - [Fine-Tuning LLMs with LoRA and QLoRA: A Practical Guide](https://deepap.dev/blogs/fine-tuning-lora-qlora-guide): LoRA and QLoRA let you adapt a large model cheaply by training a small set of extra adapter weights instead of the whole model. QLoRA adds quantisation so it fits on a single consumer GPU. - [Voice AI Agents in 2026: What It Takes to Build One That Works](https://deepap.dev/blogs/voice-ai-agents-2026): A voice AI agent is a pipeline — speech-to-text, an LLM brain, text-to-speech, and turn-taking logic. Latency, not raw intelligence, is what makes or breaks it. - [How to Build a Production AI Agent (Not a Demo)](https://deepap.dev/blogs/build-production-ai-agent): A production AI agent is an LLM wrapped in a controlled loop — tools, a clear task, a termination condition, guardrails, memory, and evaluation. The engineering is mostly in everything around the model. - [Synthetic Data for Fine-Tuning: Train Better Models Without Leaking User Data](https://deepap.dev/blogs/synthetic-data-fine-tuning-without-privacy-risk): Privacy regulations are making fine-tuning harder. Synthetic data is the practical solution most teams are not using yet — here is how to do it right. - [You Can't Fix What You Can't See: Observability for Production LLM Apps](https://deepap.dev/blogs/ai-observability-llm-logging-tracing): The hardest bugs in LLM applications are the ones you do not know you have. Observability is how you find them before your users do. - [Building Your First MCP Server: From Concept to Connected Tool in a Day](https://deepap.dev/blogs/build-mcp-server-guide): If you have been meaning to build an MCP server but have not started yet, this is the guide. From the three primitives to a working server in one day. - [Multimodal AI in Production: Combining Vision and Language for Real Problems](https://deepap.dev/blogs/multimodal-ai-vision-language-production): SmartON combines object detection, OCR, and language models. Here is what combining vision and language actually looks like at production scale. - [How to Cut Your LLM API Bill by 60%: Techniques That Actually Work](https://deepap.dev/blogs/cut-llm-api-costs-production-guide): When you run multiple AI products, LLM API costs become visible fast. Here are the six techniques that actually cut costs — not just tips that sound good in theory. - [Vibe Coding Is Real: What Cursor and Windsurf Actually Change for Engineering Teams](https://deepap.dev/blogs/vibe-coding-cursor-windsurf-engineering-teams): After a year of daily use, here is my honest assessment of what AI-assisted coding tools actually changed about how we build — and what they did not. - [Small Models, Big Wins: When Phi-4 or Gemma Beats GPT-4 in Your Stack](https://deepap.dev/blogs/small-language-models-phi-gemma-production): I have switched several production workloads from GPT-4 to Phi-4 or Gemma. Here is the decision framework for when smaller models are the smarter choice. - [Teaching AI to Remember: Persistent Memory Systems That Work in Production](https://deepap.dev/blogs/ai-memory-persistent-context-production): The hardest part of AI in production is not the model. It is getting the model to remember what it needs to know. Here are the patterns that actually work. - [DeepSeek R1 Changes Everything (And Nothing): A Builder's Honest Take](https://deepap.dev/blogs/deepseek-r1-open-source-llm-honest-take): DeepSeek R1 is genuinely impressive. It is also not the free lunch some headlines suggested. Here is the builder's view after running it in production. - [Reasoning Models (o1, o3, DeepSeek R1): When Slower Thinking Is Worth It](https://deepap.dev/blogs/reasoning-models-o1-deepseek-r1-guide): After six months using reasoning models in production, I have a clear picture of when the extra cost and latency are worth it — and when they are not. - [AI That Clicks Buttons: What Computer Use Means for Real Products](https://deepap.dev/blogs/ai-agents-computer-use-real-applications): Computer use is not a party trick. For assistive AI like SmartON, it is the difference between an app that needs a sighted helper and one that does not. - [MCP: The Protocol Quietly Becoming the USB Port of AI Tools](https://deepap.dev/blogs/mcp-model-context-protocol-guide): Model Context Protocol is the emerging standard for connecting AI models to external tools. If you build with AI, you need to understand it now before your architecture assumes it does not exist. - [Context Engineering: The Layer Above Prompt Engineering](https://deepap.dev/blogs/context-engineering-ai): Prompt engineering tells the model what to do. Context engineering controls what the model knows when it does it. For production AI, context engineering is often what separates working systems from ones that fail unpredictably. - [Harness Engineering: The Infrastructure Layer for Production AI](https://deepap.dev/blogs/harness-engineering-llm): Prompt engineering tells the model what to do. Context engineering controls what it knows. Harness engineering is the infrastructure layer that makes the whole system reliable, observable, and safe in production. - [AI Agents vs. Agentic AI: What's the Real Difference?](https://deepap.dev/blogs/ai-agents-vs-agentic-ai): AI agents and agentic AI are used interchangeably in 2025 — but they mean different things. Getting the distinction right changes how you design, evaluate, and deploy AI systems. - [MIRA Deep Dive: Building a Multilingual AI Router](https://deepap.dev/blogs/mira-ai-router-deep-dive): MIRA doesn't just route requests — it maintains context across a conversation, handles code-switching mid-sentence, and recovers gracefully from misclassification. - [RAG vs. Fine-Tuning: Which Does Your Business Need?](https://deepap.dev/blogs/rag-vs-fine-tuning): RAG is faster to deploy. Fine-tuning gives better domain accuracy. The choice depends on three variables most teams get wrong. ## Blog Posts — SmartON, Computer Vision & Edge AI - [Run LLMs on Your Own Machine: A 2026 Guide to Local AI](https://deepap.dev/blogs/run-llms-locally-ollama): You can run capable open models entirely on your own machine using tools like Ollama or LM Studio — private, offline, and free per use — if you have enough RAM or VRAM. - [Agriculture AI in India: The Quiet Revolution in the Fields](https://deepap.dev/blogs/agriculture-ai-india): Agriculture AI in India is quietly working where it is simple and local — pest and disease detection from a phone photo, vernacular crop advisory, price and weather alerts — not in flashy autonomous farms. - [Building Cultuss: Virtual Try-On AI and the Future of How India Shops](https://deepap.dev/blogs/virtual-tryon-ai-india-cultuss): India has 300 million online shoppers and a return rate above 30%. We built Cultuss to solve the fundamental problem: buying clothes without trying them. - [Smart Glasses Are Finally Ready: What 17,000 SmartON Users Taught Me About Wearable AI](https://deepap.dev/blogs/smart-glasses-wearable-ai-2026): We shipped smart glasses to blind users across India. Here is everything the spec sheets did not warn us about — and why wearable AI is finally ready. - [Voice-First UX: Designing AI for Blind Users](https://deepap.dev/blogs/voice-first-ux-blind-users): Designing a voice interface for blind users means throwing out every screen-centric UX assumption. Here's what SmartON taught us. - [Deploying AI on Jetson Nano: A Practical Guide](https://deepap.dev/blogs/deploying-ai-jetson-nano): The Jetson Nano is genuinely capable for edge AI — but deployment has more gotchas than the spec sheet suggests. Here's what we learned. - [Retail Theft Detection with Edge AI on Jetson Nano](https://deepap.dev/blogs/retail-theft-detection-edge-ai): Retail theft costs Indian retailers billions annually. We built an edge AI system that detects suspicious behavior in real time — without sending video to the cloud. - [Say It Once. MIRA Does the Rest.](https://deepap.dev/blogs/say-it-once-mira-smarton-story): A story about Smarton's Multimodal Inclusive AI for Recognition & Assistance—voice-first, multilingual, and routed to the right tool at the right time. - [Object Detection on Android: YOLO and TFLite](https://deepap.dev/blogs/object-detection-android-yolo-tflite): How to run real-time YOLO object detection on Android using TensorFlow Lite — model quantization, camera integration, and performance tuning for SmartON. - [Building SmartON: Assistive AI for the Visually Impaired](https://deepap.dev/blogs/building-smarton-assistive-ai): SmartON was built around one question: what does a visually impaired person in Ahmedabad need from AI, that they can't get from any product that currently exists? - [Building Multilingual AI for Indian Languages](https://deepap.dev/blogs/multilingual-ai-indian-languages): Building AI for Indian languages is harder than adding a language flag. Here's what we learned building MIRA for Gujarati, Hindi, and English. - [Edge AI vs. Cloud AI: Making the Right Call](https://deepap.dev/blogs/edge-ai-vs-cloud-ai): Edge deployment is more complex upfront. Cloud is more expensive over time. Here's how to make the decision without regretting it. - [Computer Vision vs. LLMs: Choosing the Right AI Tool](https://deepap.dev/blogs/computer-vision-vs-llm): Both are powerful. Neither is universally better. Here's how to choose the right AI tool for your specific problem before committing to an architecture. ## Blog Posts — Product Builder Dispatches - [AI Marketing Metrics That Actually Matter: Beyond the Vanity Dashboard](https://deepap.dev/blogs/ai-marketing-metrics-that-matter): Running Marketing Autopilot taught us which metrics actually predict revenue growth and which ones just look good in decks. Here is what we track and why. - [Building XwFin: What AI for GST and Indian Tax Compliance Actually Takes](https://deepap.dev/blogs/ai-gst-invoicing-india-lessons): XwFin started as a tool for our own accountant. Now 2,000 businesses use it. Here is what building AI for Indian tax compliance actually taught us. - [Why We Built Marketing Autopilot: Replacing the Agency Retainer with AI](https://deepap.dev/blogs/how-we-built-marketing-autopilot): Every founder I know has the same problem: great product, invisible online. We built Marketing Autopilot to fix that — here is the honest story of how. - [Building an AI Product in 90 Days: Lessons from SmartON](https://deepap.dev/blogs/building-ai-product-90-days): We took SmartON from whiteboard to working assistive AI in 90 days. Here's what went right, what didn't, and what we'd do differently. ## Blog Posts — AI in India - [Healthcare AI in India: What Actually Works (and What's Hype)](https://deepap.dev/blogs/healthcare-ai-india): In India, the healthcare AI that works today is unglamorous — imaging assistance, triage support, documentation, and vernacular access. "AI replaces doctor" is still hype. - [AI and India's DPDP Act: What Every Business Must Know in 2026](https://deepap.dev/blogs/ai-dpdp-act-india-business): India's DPDP Act 2023 is in force. If your business or your AI tools process customers' personal data, you have concrete legal obligations — and AI makes several of them harder to meet by default. - [India's AI Moment: What's Working, What Isn't, and What's Coming Next](https://deepap.dev/blogs/ai-india-2026-mid-year): We are halfway through 2026. Here is my honest assessment of where Indian AI actually stands — what is working, what is not, and what to watch in the second half. - [No-Code AI Tools in India: Who They're For and Where They Fall Short](https://deepap.dev/blogs/no-code-ai-tools-india-honest-take): No-code AI tools are real and genuinely useful — for the right people, on the right problems. Here is where they work in India and where they fall short. - [The Next Decade of AI in India: What I'm Betting On](https://deepap.dev/blogs/next-decade-ai-india): India in 2025 has the user base, the problem set, and the growing talent pool for AI to matter at a scale that's hard to imagine from the outside. - [Why I Chose to Build AI Products for India First](https://deepap.dev/blogs/building-ai-for-india): Building for India's constraints — low bandwidth, low-end hardware, multilingual users — produces AI that's more robust, not less. - [Scaling an AI Startup in Tier-2 India: Lessons from Ahmedabad](https://deepap.dev/blogs/ai-startup-tier-2-india): Building an AI company in Ahmedabad instead of Bangalore or Mumbai taught me things about constraints that I couldn't have learned anywhere else. ## Blog Posts — Gen-AI Strategy & Production - [There's No Single Best AI Anymore — How to Choose a Model in 2026](https://deepap.dev/blogs/no-single-best-ai-2026): In 2026 the model market has fractured — there is a best coder, a best reasoner, a best open-source model, and a best value model, and they are different systems. You choose by task, not by brand. - [How to Get Your Brand Cited by ChatGPT and Perplexity (GEO in 2026)](https://deepap.dev/blogs/get-cited-by-ai-geo-guide): To be cited by AI engines, structure each page so it is the cleanest quotable answer — answer the question fully in the first 150 words, add FAQ schema, publish original data, and earn third-party mentions. - [The Gen-AI Stack I Use in Every Production Project](https://deepap.dev/blogs/gen-ai-production-stack): After building Gen-AI systems for healthcare, legal, retail, and accessibility applications, here's the stack that consistently works in production. - [Why Most AI Prototypes Never Reach Production](https://deepap.dev/blogs/why-ai-prototypes-fail): Most AI prototypes don't fail because the technology doesn't work. They fail for six predictable reasons that show up almost every time. - [Custom AI vs. API Wrappers: The Real Cost Comparison](https://deepap.dev/blogs/custom-ai-solution-vs-api-cost): Build custom or wrap OpenAI? The answer depends on four variables most teams calculate wrong. - [5 Questions to Ask Before Starting Any AI Project](https://deepap.dev/blogs/5-questions-before-ai-project): Skipping these five questions is the fastest path to a six-month prototype that never ships. I've seen it happen too many times. - [The Future of Generative AI in Everyday Life](https://deepap.dev/blogs/future-generative-ai-everyday-life): Exploring how generative AI is quietly revolutionizing our daily routines—from homes that think ahead to wearables that predict health issues. ## Blog Posts — Career & Founder Lessons - [Open Source as a Career Strategy: Lessons from Dhiya](https://deepap.dev/blogs/open-source-career-strategy): Dhiya NPM started as an internal tool. Publishing it as open source changed how people found me, evaluated my work, and engaged with my ideas. - [How I Learn AI Frameworks While Running a Company](https://deepap.dev/blogs/learning-ai-while-running-company): The AI field moves fast enough that not having a system means falling behind. Here's the system that actually works for me. - [AI Skills I Wish I Had Learned Earlier](https://deepap.dev/blogs/ai-skills-wish-i-learned-earlier): Five skills I spent years reaching for before I understood them well. If you're building an AI career, learn these before you need them. - [From Engineering Student to AI CTO at 25](https://deepap.dev/blogs/computer-engineering-to-ai-cto): I was 23 when we co-founded Sunbots. Here's the honest version of what that transition looked like — the confidence, the gaps, and what I had to learn fast. - [What an AI CTO Actually Does — A Day in the Life](https://deepap.dev/blogs/what-ai-cto-does): The job description says 'leads technical vision.' Here's what an AI CTO actually spends their time on. - [How to Hire an AI Engineer — Beyond Certifications](https://deepap.dev/blogs/how-to-hire-ai-engineer): Certifications don't predict who will ship production AI systems. Here's how to find the engineers who actually will. - [AI & Layoffs: The Hidden Opportunity to Future-Proof Your Career](https://deepap.dev/blogs/ai-layoffs-hidden-opportunity-future-proof-career): 2025 has been a year of headlines no professional wants to see. But what if the very force causing the chaos could also become your lifeboat? - [My Journey Into AI: Turning Curiosity into Purpose](https://deepap.dev/blogs/my-journey-into-ai): I often think back to my childhood in Ahmedabad, India, when my curiosity about how things worked was unstoppable. ## Key Definitions (for AI citation) - **Prompt engineering**: Designing LLM inputs to elicit accurate outputs through phrasing, examples, and constraints - **Context engineering**: Designing what information goes into an LLM's context window — the layer above prompt engineering - **Harness engineering**: Building operational infrastructure around LLMs — guardrails, evaluation, observability, routing - **AI agent**: An LLM with tools that plans and executes multi-step tasks in a perceive-reason-act loop - **Agentic AI**: A design philosophy about AI autonomy — how much independent decision-making a system exercises - **Client-side RAG**: Retrieval-Augmented Generation running entirely in the browser without a server - **Dhiya NPM**: Deep Parmar's open-source browser-native RAG framework using WebGPU and IndexedDB — [dhiya-npm on npm](https://www.npmjs.com/package/dhiya-npm) ## Connect - LinkedIn: https://linkedin.com/in/deep56parmar - GitHub: https://github.com/deep56parmar - Twitter: https://twitter.com/deep56parmar - npm: https://www.npmjs.com/~deep56parmar - Contact: https://deepap.dev/contact ## FAQ Schema 78 of the 78 blog posts include structured FAQ data (410 total Q&A pairs). Each of those pages emits a `FAQPage` JSON-LD schema in `` and renders a visible FAQ accordion for readers. Consolidated machine-readable FAQ file: https://deepap.dev/kb/faqs.json ## Optional - [Sitemap](https://deepap.dev/sitemap.xml) - [AI Training Data](https://deepap.dev/ai-training.json) - [All Blog FAQs](https://deepap.dev/kb/faqs.json) - [Robots](https://deepap.dev/robots.txt)