# Divyansh Shukla AI engineer building agents, automation and the products around them. India. **Status:** Open to AI engineering roles, and to conversations with investors and partners about these products. Based in India; remote works. **Contact:** shukladivyansh953@gmail.com · [GitHub](https://github.com/TheDivyanshShukla) · [LinkedIn](https://www.linkedin.com/in/thedivyanshshukla) ## Abstract I build AI products end to end and ship them. This year that meant Ice, a desktop assistant that does the work in your apps; Blank, a website builder that hands back real code; Synapse, a chatbot that answers a site’s visitors; and CampusQuiz, which turns class notes into a live quiz. Underneath them is research: in Plan Once, Ground Locally [[3]](https://doi.org/10.5281/zenodo.22904595), I show that a web agent which writes its whole plan in one LLM call, then grounds each step with a small on-device model, matches a budget-matched ReAct baseline on success while using 83.6% fewer input tokens. Most of the products ship under NARA [[4]](https://naravirtual.ai). This page is written for agents as well as people: it serves an MCP endpoint, an llms.txt and a markdown copy of every page (see Agent access below). ## Selected work - [Plan Once, Ground Locally](https://divyanshshukla.com/research/plan-once-ground-locally) (Research, 2026): Preprint and 200-task benchmark: plan-and-execute web agents match budget-matched ReAct on success with 83.6% fewer input tokens. - [Ice](https://divyanshshukla.com/work/ice) (Product, 2026): An AI desktop assistant with a face: it listens, then works in your browser, files and apps, asking before anything you haven’t allowed. - [Blank](https://divyanshshukla.com/work/blank) (Product, 2026): An AI website builder that hands back the project: real React files, a database in your own Supabase account, your own domain and a push to your own GitHub. - [Synapse](https://divyanshshukla.com/work/synapse) (Product, 2026): An AI chatbot added to any website with one script tag: it learns your pages, PDFs and FAQs, answers visitors around the clock, captures leads and hands off to a person. - [CampusQuiz](https://divyanshshukla.com/work/campusquiz) (Product, 2026): AI quizzes for the classroom: a teacher uploads notes or a photo of the board, gets a quiz, and runs it live while students join with a code. - [CampusPrep.in](https://divyanshshukla.com/work/campusprep) (Product, 2026): Free RGPV syllabus, previous-year papers and unit-wise short notes for 570 subjects, first year to eighth semester, readable without an account. - [Saakshya](https://divyanshshukla.com/work/saakshya) (Hackathon, 2026): AI and blockchain document verification that returns one of four honest verdicts, from PROVEN to REJECTED, with a credential that verifies offline. Everything else: https://divyanshshukla.com/work ## Research **Plan Once, Ground Locally: Plan-and-Execute Web Agents Match ReAct Accuracy at One-Sixth of the Token Cost.** Divyansh Shukla. Preprint, Zenodo, 2026. DOI 10.5281/zenodo.22904595. Against the budget-matched baseline, planning once is not significantly different on success (88.8% vs 92.0%; paired difference −3.2 points, 95% CI −7.7 to +1.2) while using 83.6% fewer input tokens, 67.3% fewer LLM calls and 80.4% less money, and finishing 4.1× faster. Details: https://divyanshshukla.com/research/plan-once-ground-locally · PDF: https://zenodo.org/records/22904595/files/plan-once-ground-locally.pdf ## Notes - [What 2,800 browser-agent runs taught me about planning](https://divyanshshukla.com/notes/what-2800-browser-agent-runs-taught-me) (2026-10-03): Plan-and-execute web agents matched a budget-matched ReAct baseline on 200 tasks with 83.6% fewer input tokens. What held up, what didn’t, and what to copy. - [Making my site readable by agents](https://divyanshshukla.com/notes/making-my-site-readable-by-agents) (2026-10-03): This site serves an MCP server, WebMCP tools, llms.txt and a markdown copy of every page, all from one content source. What each piece does and how to try it. ## Agent access - MCP (Streamable HTTP): https://divyanshshukla.com/mcp — e.g. `claude mcp add --transport http divyansh https://divyanshshukla.com/mcp` - llms.txt: https://divyanshshukla.com/llms.txt and https://divyanshshukla.com/llms-full.txt - Markdown: append .md to any page URL, or send `Accept: text/markdown` ## Background Divyansh Shukla is an AI engineer in India who builds and ships AI products: Ice, Blank, Synapse, Conference and Loci at NARA; CampusQuiz and CampusPrep.in for students and teachers; and Saakshya, an AI and blockchain document-verification system built at the SISTec Innovation Hackathon 2026. Author of Plan Once, Ground Locally (2026). Research interests: planning and grounding for web agents, inference cost, on-device decision models, and agent tooling such as MCP. Education: B.Tech, Computer Science (AI & Data Science). - Languages: Python, TypeScript, JavaScript, SQL - Agents and LLMs: Anthropic and OpenAI APIs, OpenRouter, MCP, Playwright, LangChain, RAG with ChromaDB and Pinecone - Models on device: MLX, ModernBERT, MediaPipe, OpenCV - Web: SvelteKit, Next.js, FastAPI, Django, Tailwind CSS - Data and infrastructure: Postgres, Redis, Docker, Dokploy, Vercel, Cloudflare - Security and hardware: Aircrack-ng, Hashcat, ESP32, Raspberry Pi ## References 1. [Divyansh Shukla. GitHub profile, TheDivyanshShukla.](https://github.com/TheDivyanshShukla) 2. [Divyansh Shukla. LinkedIn profile.](https://www.linkedin.com/in/thedivyanshshukla) 3. [D. Shukla. Plan Once, Ground Locally: Plan-and-Execute Web Agents Match ReAct Accuracy at One-Sixth of the Token Cost. Preprint, Zenodo, 2026. doi:10.5281/zenodo.22904595.](https://doi.org/10.5281/zenodo.22904595) 4. [NARA, the focused software house. naravirtual.ai.](https://naravirtual.ai) 5. [Saakshya: verify the claim, not just the paper. Live system.](https://saakshya.divyanshshukla.com) 6. [Ice: the AI desktop assistant that does the work. ice.naravirtual.ai.](https://ice.naravirtual.ai) Last updated 2026-10-03. --- # Work — Divyansh Shukla Everything shipped: research, products, open source, client sites and earlier projects. ## Research - [Plan Once, Ground Locally](https://divyanshshukla.com/research/plan-once-ground-locally) (Research, 2026): Preprint and 200-task benchmark: plan-and-execute web agents match budget-matched ReAct on success with 83.6% fewer input tokens. ## Product - [Ice](https://divyanshshukla.com/work/ice) (Product, 2026): An AI desktop assistant with a face: it listens, then works in your browser, files and apps, asking before anything you haven’t allowed. - [Blank](https://divyanshshukla.com/work/blank) (Product, 2026): An AI website builder that hands back the project: real React files, a database in your own Supabase account, your own domain and a push to your own GitHub. - [Synapse](https://divyanshshukla.com/work/synapse) (Product, 2026): An AI chatbot added to any website with one script tag: it learns your pages, PDFs and FAQs, answers visitors around the clock, captures leads and hands off to a person. - [CampusQuiz](https://divyanshshukla.com/work/campusquiz) (Product, 2026): AI quizzes for the classroom: a teacher uploads notes or a photo of the board, gets a quiz, and runs it live while students join with a code. - [CampusPrep.in](https://divyanshshukla.com/work/campusprep) (Product, 2026): Free RGPV syllabus, previous-year papers and unit-wise short notes for 570 subjects, first year to eighth semester, readable without an account. - [Conference](https://divyanshshukla.com/work/conference) (Product, 2026): Browser video meetings for up to 1,000 people and webinars to 5,000, with cloud recording, scheduling and self-serve billing. - [Loci](https://loci.naravirtual.ai) (Product, 2026): Speed reading meets the method of loci: RSVP reading, AI-generated memory-palace imagery for each paragraph, and quizzes that check what stuck. - [NARA](https://naravirtual.ai) (Product, 2026): The focused software house behind Ice, Blank, Synapse, Conference and Loci, which also builds websites, chatbots and automation for clients. - [SkinDost](https://skindost.in) (Product, 2026): Online dermatology for India: a skin scan and report, video consultations with certified dermatologists, digital prescriptions and medicine delivery. ## Hackathon - [Saakshya](https://divyanshshukla.com/work/saakshya) (Hackathon, 2026): AI and blockchain document verification that returns one of four honest verdicts, from PROVEN to REJECTED, with a credential that verifies offline. ## Open source - [atlas-mcp](https://divyanshshukla.com/work/atlas-mcp) (Open source, 2026): An MCP server and CLI that connects coding agents to a team knowledge hub: agents pull approved prompts, docs and scoped secrets, and their work is captured. - [claude-code-motivator](https://divyanshshukla.com/work/claude-code-motivator) (Open source, 2026): A plain-text directive API for autonomous coding agents: polled before the agent stops, it returns one precise instruction to find, verify and ship the next task. ## Client - [The Shining 32 Dental Clinic](https://theshining32.com) (Client, 2026): Website for a two-clinic dental practice in Indore, with an admin CMS so the clinic edits its own pages. - [Café Latté](https://cafelatte.divyanshshukla.com) (Client, 2026): Website and table-side app for a lakeside café and patisserie at Noor-Us-Sabah Palace, Bhopal: menu, seating, directions, installable and usable offline. - [Aawaz Academy of Music](https://aawazacademy.com) (Client, 2026): Website, member area and staff console for a music school in Navlakha, Indore. ## Earlier - [WPA2 handshake research](https://github.com/TheDivyanshShukla/wifi-crack) (Earlier, 2025): Wireless security research: capture a WPA2 handshake and crack it offline to show why weak passphrases fail. For networks you own or may test. - [JARVIS](https://github.com/TheDivyanshShukla/jarvis-69) (Earlier, 2024): A voice-controlled assistant inspired by Iron Man’s: speech recognition, spoken replies and control of the computer. - Gesture control (Earlier, 2024): A computer-vision virtual mouse: move the cursor and drive media apps with hand gestures tracked in real time. ## Archived - SCANCROS (Archived, 2024–25): A toolkit for bug-bounty hunters that automated vulnerability-scanning pipelines. The site is retired. --- # Plan Once, Ground Locally: Plan-and-Execute Web Agents Match ReAct Accuracy at One-Sixth of the Token Cost Divyansh Shukla · Preprint · Zenodo · 2026-09-23 · DOI [10.5281/zenodo.22904595](https://doi.org/10.5281/zenodo.22904595) · CC BY 4.0 ## Abstract LLM web agents usually follow the ReAct pattern: the model is called again after every browser action, so each click is paid for with a network round trip and with a prompt that has grown since the last one. We study the alternative in which the model writes the whole plan in one call and the steps are executed locally, with a small on-device decision model (Laya, a ModernBERT-large “System-1” model used zero-shot) mapping each step to a page element, judging when asynchronous content has arrived, and checking actions that left the page unchanged. We evaluate on a new benchmark of 200 tasks across seven public websites whose ground truth is scraped from the sites themselves or produced by scripted browser runs, and graded by fixed regular expressions. Every task was run three times under four configurations with the same LLM (DeepSeek v4.1 flash): ReAct, ReAct with a reading and turn budget matched to the plan agent, plan-and-execute with Laya, and an ablation that replaces Laya with word-overlap matching, for 2,400 runs, plus 400 further runs with a second model. Against the budget-matched baseline, planning once is not significantly different on success (88.8% vs 92.0%; paired difference −3.2 points, 95% CI −7.7 to +1.2) while using 83.6% fewer input tokens, 67.3% fewer LLM calls and 80.4% less money, and finishing 4.1× faster. Two negative results accompany that headline. Against the unmatched baseline the plan agent also looks more accurate (+8.2 points, p < 0.001), but the entire gap is an artefact of the baseline’s smaller reading window and turn cap. And the neural executor is not significantly better than word-overlap matching (+2.3 points, 95% CI −0.2 to +5.0). A second, cheaper model reproduces the token and call savings and widens the accuracy gap in the plan agent’s favour, but not the money saving: in 7.5% of its runs it looped while writing the plan and hit the provider’s generation cap. ## Results 200 tasks × 3 repeats per configuration, DeepSeek v4.1 flash at temperature 0. Means per task. | Configuration | Success (95% CI) | Time | LLM calls | Input tokens | Cost | |---|---|---|---|---|---| | ReAct | 80.7% (77.3–83.6) | 32.8 s | 6.03 | 20,523 | $0.00186 | | ReAct, budget-matched | 92.0% (89.6–93.9) | 42.2 s | 6.22 | 24,697 | $0.00202 | | Plan + Laya | 88.8% (86.1–91.1) | 10.2 s | 2.03 | 4,060 | $0.00040 | | Plan + word overlap | 86.5% (83.5–89.0) | 10.8 s | 2.12 | 4,305 | $0.00042 | ## Findings - **Accuracy is a tie once budgets match.** Plan + Laya vs budget-matched ReAct: −3.2 points [−7.7, +1.2], not significant in any repeat (McNemar p = 0.31, 0.54, 0.12). The apparent win over plain ReAct comes from its smaller extract window and turn cap; raising both gains ReAct 11.3 points. - **Efficiency is the real result.** 83.6% fewer input tokens, 67.3% fewer LLM calls, 80.4% less money and 4.1× less wall-clock time (Wilcoxon p < 1e-25 each). On the 505 paired runs both solved: 3,905 vs 13,298 tokens, 1.88 vs 4.41 calls, 9.6 vs 17.4 s. - **The on-device grounder is not a significant win.** Laya vs word matching: +2.3 points [−0.2, +5.0], p = 0.07. It helps where identical controls need context (saucedemo: 60/60 vs 49/60 runs). Grounding calls take 31 ms median. - **Where planning ahead fails.** TodoMVC: 0/30 and 2/30 runs for the plan configurations vs 26/30 for budget-matched ReAct. A plan written before the page exists cannot name list items that appear only after typing them. ## What the record contains - The 200 tasks with ground truth and sources - One graded row per run for all 2,800 runs - Full per-run traces: every prompt, completion, tool result and model decision - The agent code and the analysis scripts that recompute every number, table and figure ## Cite ```bibtex @misc{shukla2026planonce, title = {Plan Once, Ground Locally: Plan-and-Execute Web Agents Match ReAct Accuracy at One-Sixth of the Token Cost}, author = {Shukla, Divyansh}, year = {2026}, month = sep, publisher = {Zenodo}, doi = {10.5281/zenodo.22904595}, url = {https://doi.org/10.5281/zenodo.22904595}, note = {Preprint} } ``` ## Links - PDF: https://zenodo.org/records/22904595/files/plan-once-ground-locally.pdf - Zenodo record: https://zenodo.org/records/22904595 - Code and data: https://github.com/TheDivyanshShukla/plan-once-ground-locally --- # Notes — Divyansh Shukla Short working papers on agents, inference cost and shipping AI products. - [What 2,800 browser-agent runs taught me about planning](https://divyanshshukla.com/notes/what-2800-browser-agent-runs-taught-me) (2026-10-03): Plan-and-execute web agents matched a budget-matched ReAct baseline on 200 tasks with 83.6% fewer input tokens. What held up, what didn’t, and what to copy. - [Making my site readable by agents](https://divyanshshukla.com/notes/making-my-site-readable-by-agents) (2026-10-03): This site serves an MCP server, WebMCP tools, llms.txt and a markdown copy of every page, all from one content source. What each piece does and how to try it. --- # What 2,800 browser-agent runs taught me about planning 2026-10-03 · web agents, ReAct, plan-and-execute, benchmarks, inference cost > Plan-and-execute web agents matched a budget-matched ReAct baseline on 200 tasks with 83.6% fewer input tokens. What held up, what didn’t, and what to copy. Most LLM web agents follow ReAct: the model picks one action, the browser runs it, and the model is called again with everything so far. Every click costs a network round trip and a prompt that has grown since the last one. I wanted to know what happens if the model writes the whole plan once and the steps are carried out locally. So I built both, and a benchmark to compare them fairly. The full write-up is the preprint [Plan Once, Ground Locally](https://doi.org/10.5281/zenodo.22904595); this note is the short version, with the parts I would tell another engineer first. ## The setup - **200 tasks on seven public websites**, each with ground truth scraped from the site itself or produced by a scripted browser run, and graded by a fixed regular expression. No model grades another model. - **Four configurations**, three repeats each, all on DeepSeek v4.1 flash at temperature 0: ReAct; ReAct with the same reading window and turn cap as the plan agent; plan-and-execute with Laya; and plan-and-execute with plain word overlap in place of Laya. - **Laya** is a ModernBERT-large decision model used zero-shot as a small “System 1”. It maps each planned step to a page element, judges when asynchronous content has arrived, and checks actions that changed nothing visible. It runs on the device, at a median of 31 ms per call. - **2,800 recorded runs**: 2,400 on the main model and 400 more on a second, cheaper one. Every prompt, completion, tool result and model decision is in the published record. ## Accuracy is a tie once budgets match Against the budget-matched ReAct baseline, planning once is not significantly different on success: 88.8% against 92.0%, a paired difference of −3.2 points with a 95% interval from −7.7 to +1.2. The first comparison I ran looked much better, and it was wrong. Against plain ReAct the plan agent appeared 8.2 points more accurate (p < 0.001), but the whole gap came from plain ReAct’s smaller reading window and turn cap. Raising both gained ReAct 11.3 points. If your baseline is starved, your method will look brilliant. ## The savings are the real result | Configuration | Success | Time | LLM calls | Input tokens | |---|---|---|---|---| | ReAct | 80.7% | 32.8 s | 6.03 | 20,523 | | ReAct, budget-matched | 92.0% | 42.2 s | 6.22 | 24,697 | | Plan + Laya | 88.8% | 10.2 s | 2.03 | 4,060 | | Plan + word overlap | 86.5% | 10.8 s | 2.12 | 4,305 | _Means per task over 200 tasks and three repeats, DeepSeek v4.1 flash._ Against the budget-matched baseline, the plan agent used 83.6% fewer input tokens, 67.3% fewer LLM calls and 80.4% less money, and finished 4.1 times faster. On the 505 paired runs that both solved, it was 3,905 against 13,298 tokens, 1.88 against 4.41 calls, and 9.6 against 17.4 seconds. ## The clever part mattered less than I hoped Laya was not significantly better than word overlap overall: +2.3 points, with an interval from −0.2 to +5.0 (p = 0.07). It earns its place where several controls look identical and context decides, as on saucedemo, where it solved 60 of 60 runs against 49 of 60. Elsewhere, matching words got most of the way. ## Where planning ahead breaks On TodoMVC the plan configurations solved 0 and 2 of 30 runs, against 26 of 30 for budget-matched ReAct. A plan written before the page exists cannot name list items that only appear after you type them. Pages that build themselves as you act need the model back in the loop. The cheaper second model reproduced the token and call savings and widened the accuracy gap in the plan agent’s favour, but not the money saving: in 7.5% of its runs it looped while writing the plan and hit the provider’s generation cap. ## What I would copy - Default to one plan, and call the model again only to replan after a failure and to write the answer: about two calls per task instead of six. - Ground steps locally with something cheap, and check that the clever version beats word overlap before paying for it. - Budget-match your baselines. A starved ReAct makes any alternative look good. - Keep a fallback for pages that only exist after you act on them. - Publish the traces, so every number can be recomputed from the recorded runs. The paper, the 200 tasks with their ground truth, all 2,800 graded runs and the analysis scripts are on [Zenodo](https://zenodo.org/records/22904595), and the agent code is on [GitHub](https://github.com/TheDivyanshShukla/plan-once-ground-locally). --- # Making my site readable by agents 2026-10-03 · MCP, WebMCP, llms.txt, structured data, SEO > This site serves an MCP server, WebMCP tools, llms.txt and a markdown copy of every page, all from one content source. What each piece does and how to try it. More and more people meet a developer through an agent rather than a browser tab. Someone asks Claude, ChatGPT or Perplexity who you are and what you have built, and the agent reads your site on their behalf. So I rebuilt this site to be read by agents as carefully as by people, and to give them the same answer the page gives. ## One source, five surfaces Everything the site says lives in a few typed content files: a profile, the list of work, the paper and these notes. The HTML pages, the markdown versions, llms.txt, the structured data and the MCP tools are all generated from them, so an agent never reads a different story from the one on the page. ## MCP: ask the site directly There is a read-only MCP server at `/mcp`, over Streamable HTTP. Add it to Claude Code with one command: ```sh claude mcp add --transport http divyansh https://divyanshshukla.com/mcp ``` It has six tools: `about`, `list_work`, `get_work`, `get_paper`, `search` and `contact`. Each returns markdown, every input is validated, and every tool is marked read-only, so a client can call them without stopping to ask. ## WebMCP: the same tools inside the browser Agents that run inside the browser can use the same tools on the page itself, registered through WebMCP, with nothing to install. ## Markdown and llms.txt Append `.md` to any URL, or send `Accept: text/markdown`, and you get the page as clean markdown with a canonical link back to the HTML version. [/llms.txt](https://divyanshshukla.com/llms.txt) lists every page with a one-line description, and [/llms-full.txt](https://divyanshshukla.com/llms-full.txt) is the whole site in one file, which is often all an agent needs. ```sh curl -H "Accept: text/markdown" https://divyanshshukla.com/ ``` ## Structured data for search and answer engines Every page carries schema.org JSON-LD that points to a single `Person` node, so search engines and answer engines resolve one entity for my name. Products are marked as software applications, the paper as a scholarly article with Google Scholar tags, and these notes as blog posts. `robots.txt` welcomes AI crawlers: being cited correctly is the point. ## The page shows its own grounding Figure 1 on the [home page](https://divyanshshukla.com/) is an agent that runs in your browser. Ask it something and it plans once, outlines the evidence on the page as it finds it, and answers, with no model call and no tokens spent. It is the idea from my paper, made small enough to run on a portfolio. ## Try it - Connect the MCP server and ask your agent what I have shipped. - Open [llms.txt](https://divyanshshukla.com/llms.txt), or any page with `.md` on the end. - Switch on **Agent view** in the toolbar to outline every element an agent can act on. --- # Ice Product · 2026 · NARA > An AI desktop assistant with a face: it listens, then works in your browser, files and apps, asking before anything you haven’t allowed. Ice sits on the desktop as a small character you can move. Talk to it and it works across the browser, files and apps, leaving a trail of every step: what it opened, what it wrote, what failed and how long it took. The macOS beta comes first, Windows next. It is not a chat box. You watch Ice work the way you would watch a colleague, and it asks before doing anything you haven’t allowed. Reservations are open at $5, refundable until launch, and the first 100 paid reservations become founding members who keep the desktop app for life. ## Highlights - Talk to it, and it acts in your browser, your files and your apps. - Every step lands on a trail you can read, with timings and failures. - Anything you haven’t allowed waits for your approval. - macOS beta first, then Windows; $5 reservations, refundable until launch. ## Links - Live: https://ice.naravirtual.ai - Page: https://divyanshshukla.com/work/ice --- # Blank Product · 2026 · NARA > An AI website builder that hands back the project: real React files, a database in your own Supabase account, your own domain and a push to your own GitHub. Describe a site and Blank writes a real React project you can read, edit, connect to your own database and publish on your own domain. Nothing is locked in, because there is nothing to lock: the code, the data and the repository are yours from the first build. Most AI site builders keep the result inside their editor. Blank goes the other way: every build is a project you could hand to any developer, with a Postgres database and row-level security in your own Supabase account, a custom domain, version history, and a one-click push to your own GitHub repository. ## Highlights - Real React source you can read and edit, not a locked editor. - The database lives in your own Supabase account, with row-level security. - Custom domains and a one-click push to your own GitHub repository. - Version history on every build. ## Links - Live: https://nvblank.com - Page: https://divyanshshukla.com/work/blank --- # Synapse Product · 2026 · NARA > An AI chatbot added to any website with one script tag: it learns your pages, PDFs and FAQs, answers visitors around the clock, captures leads and hands off to a person. Synapse trains on a site’s own content, answers customers at any hour, collects leads and passes the conversation to a human when it should. The free plan covers 250 messages a month without a card. It is built for businesses whose website already holds the answers: one script tag installs it, replies come back in well under a second, and when the content doesn’t cover a question, Synapse says it doesn’t know rather than guessing. ## Highlights - Installs with one script tag. - Trains on a site’s pages, PDFs and FAQs. - Captures leads and hands the conversation to a person when it should. - Free plan: 250 messages a month, no card. ## Links - Live: https://naravirtual.ai/synapse - Page: https://divyanshshukla.com/work/synapse --- # CampusQuiz Product · 2026 > AI quizzes for the classroom: a teacher uploads notes or a photo of the board, gets a quiz, and runs it live while students join with a code. A teacher uploads notes, a PDF or a photo of the board, and CampusQuiz drafts the quiz: multiple choice, true or false, or mixed. Students join with a six-character code, with nothing to install, attempt it live, and see their score and the explanations the moment they submit. While the class works, a live board shows who has joined, who has submitted, the correct rate for each question and the weakest topics so far. Across quizzes, each student’s dashboard tracks the concepts they keep getting wrong. It is free for students, built for teachers, colleges and coaching institutes in India, and the sister product of CampusPrep.in. ## Highlights - AI-drafted quizzes from notes, PDFs or a photo of the board. - Students join with a six-character code or a link; nothing to install. - A live classroom board that updates as answers come in. - Weak-topic tracking for every student across every quiz. - Anonymous mode for quick pulse checks, named mode for graded work. ## Links - Live: https://campusquiz.in - Page: https://divyanshshukla.com/work/campusquiz --- # CampusPrep.in Product · 2026 > Free RGPV syllabus, previous-year papers and unit-wise short notes for 570 subjects, first year to eighth semester, readable without an account. A study site for RGPV B.Tech students: pick a branch and semester to get the official syllabus, solved and unsolved previous-year papers, short notes and important questions, with nothing behind a sign-up wall. It covers 570 subjects from first year to eighth semester. On top of the papers sit subject analytics and a personal workspace where students keep their own notes and practice questions. CampusQuiz, its sister product, brings the same students into live class quizzes. ## Highlights - 570 subjects, first year to eighth semester. - Official syllabus, solved and unsolved previous-year papers, short notes and important questions. - Nothing behind a sign-up wall. - Subject analytics and a personal workspace for notes and practice questions. ## Links - Live: https://campusprep.in - Page: https://divyanshshukla.com/work/campusprep --- # Conference Product · 2026 · NARA > Browser video meetings for up to 1,000 people and webinars to 5,000, with cloud recording, scheduling and self-serve billing. HD meetings and webinars that run in any browser, with breakout rooms, polls and Q&A, a whiteboard, waiting rooms, screen sharing, cloud recording and admin audit logs. Conference is NARA’s browser-first alternative to Zoom: there is nothing to install, anyone can join by meeting ID, billing is self-serve, and the same room scales from a one-to-one call to an all-hands. An Android app is on Google Play. ## Highlights - HD meetings for up to 1,000 participants; webinars to 5,000 attendees. - Cloud recording, scheduling and a waiting room. - Polls and Q&A, reactions, a whiteboard, breakout rooms and screen sharing. - Admin and audit logs, and self-serve billing. ## Links - Live: https://conference.naravirtual.in - Page: https://divyanshshukla.com/work/conference --- # Saakshya Hackathon · 2026 > AI and blockchain document verification that returns one of four honest verdicts, from PROVEN to REJECTED, with a credential that verifies offline. Built with Team Saakshya at the SISTec Innovation Hackathon 2026, for a problem statement on blockchain-based document verification. Upload a marksheet, a degree or a PAN card: Saakshya reads and audits it, checks the claim against the issuing authority’s register, anchors a salted hash in a Merkle-batched ledger and returns a signed credential with a QR code that verifies offline. The point is the verdict. Instead of a yes or no, every document gets one of four levels: L3 PROVEN (issuer-signed or from DigiLocker), L2 CONFIRMED (matches the official register), L1 PLAUSIBLE (clean, but no register to check) and L0 REJECTED (evidence of tampering). ## Highlights - OCR plus five families of forensic signals: error level analysis, noise residual, copy-move detection, metadata and PDF structure, and template consistency. - Hash-chained ledger blocks with Merkle roots and RFC 9162-style inclusion proofs, using the same data model as the Hyperledger Fabric chaincode it would swap in for. - On the ledger: salted leaf hash, Merkle root, verdict, model version and revocations. Off it, encrypted at rest: the document image, extracted personal data, salts and keys. - A public verifier that checks the Ed25519 signature in the browser with WebCrypto and keeps working offline after the first load. ## Built with FastAPI, RapidOCR, SvelteKit (Svelte 5), Tailwind CSS v4, GSAP, WebCrypto, Ed25519, Merkle ledger ## Links - Live: https://saakshya.divyanshshukla.com - Source: https://github.com/TheDivyanshShukla/saakshya - Page: https://divyanshshukla.com/work/saakshya --- # atlas-mcp Open source · 2026 > An MCP server and CLI that connects coding agents to a team knowledge hub: agents pull approved prompts, docs and scoped secrets, and their work is captured. atlas-mcp connects any coding agent, including Claude Code, Cursor, Windsurf, VS Code with Copilot and Cline, to a team’s Atlas knowledge hub. The agent pulls approved prompts, documents and scoped secrets as context, and Atlas records what the agent builds, so the team can see the work without anyone updating a ticket. It is a thin, cached client over the Atlas REST API. Reads come from a local stale-while-revalidate cache, so they are instant and work offline; writes queue and flush when the network returns. Authentication, fine-grained scopes and audit stay on the server, and tools a key cannot use are hidden from the agent. ## Highlights - Toolsets gated by key scope: context (browse, read, search, get a secret), write, capture (log work, save a prompt) and tasks. - One command per repository, `atlas .`, binds the repo to a project through a `.atlas` file; the key never lives in the file. - `atlas hooks install` captures every Claude Code prompt and session with no cooperation needed from the agent. - Remote HTTP transport for zero install, or local stdio for repo-aware capture with the offline cache. ## Built with JavaScript, Bun, Model Context Protocol, stdio and Streamable HTTP, Claude Code hooks ## Links - Source: https://github.com/TheDivyanshShukla/atlas-mcp - Page: https://divyanshshukla.com/work/atlas-mcp --- # claude-code-motivator Open source · 2026 > A plain-text directive API for autonomous coding agents: polled before the agent stops, it returns one precise instruction to find, verify and ship the next task. Coding agents tend to stop at the first green checkmark. Wire this endpoint into a Stop hook and, before the agent ends its turn, it fetches one line of senior-engineer operating advice: find the next high-value task, verify the work, commit it and continue. GET / returns one newline-terminated line of text/plain. Modes steer the advice toward continuing, verifying, shipping, focus or quality; the default deliberately never says stop, and an explicit stop mode is the honest counterweight. The corpus is static and hand-curated, with a no-repeat window so consecutive calls differ, and no external calls, keys or rate limits. ## Highlights - Seven modes: continue, verify, ship, focus, quality, stop and any. - JSON output, an OpenAPI page and a live stats dashboard at /stats. - Works as a Claude Code Stop hook or as a one-line CLAUDE.md instruction. ## Built with Python, FastAPI, uv, Docker ## Links - Source: https://github.com/TheDivyanshShukla/claude-code-motivator - Page: https://divyanshshukla.com/work/claude-code-motivator