Hey, I'm Varun.

Shipped in 2026.

• · Zopeo, co-founder - customer retention for indian merchants — 300+ of them run on it. small shops here lose customers silently. no crm, no idea who stopped coming back, no way to tell a quiet tuesday apart from a customer who has gone for good. zopeo turns a qr scan at the counter into a customer record, then tells the shop who is drifting and what would bring them back. more spring boot and mysql behind a react dashboard, a react native app for customers, and an admin portal. razorpay for payments, firebase for push, google maps for discovery, redis and shedlock so schedulers cannot double-fire across instances, resilience4j on the outbound calls. the analytics side finds peaks, abandonment and anomalies; the ai layer reads those signals and drafts offers as strict json — off-peak fillers, conversion boosters, retention campaigns — for a merchant to review. it never auto-publishes, and when the data is too thin it falls back to the store catalogue instead of refusing to help.
• · CleanCar, founder - a car-cleaning business that runs entirely on whatsapp — 900+ customers, six figures a month. customers book in a chat, field workers get their jobs in the same chat, and nobody installs anything. every ops tool i looked at assumed a play store download and a user who reads english. more one spring boot app, one mysql database, no queue. inbound webhooks are acknowledged in milliseconds and processed by a poller, outbound replies go through a transactional outbox so a reply is never lost and never sent twice. worker-facing copy ships in english, hindi and telugu.
• · keelson - your repo's architecture rules, enforced the moment an agent tries to break them. i kept writing rules into CLAUDE.md and watching the agent take the shortest path to a working diff anyway. prompts are suggestions. hooks are enforcement. more one rule file, any language, works at agent edit time and in CI. the agent reads the denial and corrects itself in the same turn, so the bad commit never exists. init sets the first line limit just above your repo's 95th-percentile file, because a linter that fails four hundred times on install gets uninstalled.
• · agentrouter - reads a repo and tells you which subagents actually earn their existence. every other tool for this emits backend-agent, frontend-agent, database-agent. those are locations, not tasks — the agent could already read those directories. you end up with ten agents nobody invokes and delete all of them within a week. more a subagent only earns its existence by changing the outcome, and there are exactly five mechanisms by which it can: context isolation, tool scoping, effort tier, and two more. anything that fails all five is named out loud as a skill or a rule instead. no api key, no model call, deterministic and testable against fixtures.
• · crosscheck-ai - one coding agent writes, a different one reviews, and code controls the loop. a model reviewing its own diff approves it. so the author and the reviewer are different agents from different providers, and neither of them gets to decide when they're finished. more task, author edits, tests, reviewer reads the diff, author corrects, tests, decision. bounded correction rounds and deterministic checks between every step. agents do not choose the workflow, approve their own work, or debate forever.
• · termvoice - push-to-talk prompts for terminal ai agents. typing long prompts into a cli is the slowest part of using one. transcription runs locally with whisper, and nothing leaves the machine until i have read it back and approved it. more no account, no server, no telemetry, no api key. temporary recordings are deleted automatically, and a transcript is never evaluated as a shell command — it only ever becomes a prompt you approved.
• · termfetch - turns your github profile into a neofetch-style terminal card. every github profile header is the same three badge services. this one renders your own photo as coloured terminal art next to live stats, as a single self-contained svg committed to a repo you control. more no image service at view time, so nothing breaks when someone else shuts a free tier down. a scheduled action refreshes the card without spending api quota on every README view.
• · RealSecurePdf - rearrange and merge pdfs without the bytes ever leaving your browser. every free pdf tool asks you to upload the document first. this one has no upload endpoint at all. more imports fail closed on encryption, signatures, active content, embedded files and interactive forms, so it can never silently invalidate a signature or change how a form behaves. before download the generated pdf is reopened twice — once by pdf-lib, once by pdf.js — and checked page by page.
• · ats-resume-tailor - rewrites a resume against one job description, then measures the result. the model never grades its own work here — a deterministic scorer does, and the highest-scoring draft wins. self-graded rewrites drift; measured ones converge. more your pdf is parsed in the browser and never uploaded — only the extracted text goes to the worker. the scorer weights keyword coverage at 55, parser-critical structure at 30, one-page fit at 10, and absence of keyword stuffing at 5. a refinement that overreaches can never make the delivered resume worse than the draft before it.
• · Voix - audio-first social — posts and replies as voice instead of text. i wanted a feed that sounded like people rather than reading like a press release. built the whole thing, couldn't get it to scale, and stopped. i'd rather say that here than quietly delete the repo.
• · insurance-mcp-poc - an mcp server that hands an agent a real life-insurance back office. every mcp demo runs on a toy dataset, and the interesting problems only show up on a dense one. policy documents, in-flight applications, underwriting rules — twenty-four tools any mcp client can call. more implement the domain once as a server and every mcp-capable client gets the same tools with the same contracts, instead of a bespoke integration per assistant. all the data is synthetic.

Packages.

• · @saivarun1410/keelson - architecture rules, enforced at agent edit time npm i -D @saivarun1410/keelson 121/mo
• · @saivarun1410/agentrouter - which subagents your repo actually needs npm i -g @saivarun1410/agentrouter 49/mo
• · crosscheck-ai - author-reviewer orchestration across providers npm i -g crosscheck-ai 199/mo
• · termvoice - local push-to-talk for terminal agents pip install termvoice 129/mo
• · termfetch - neofetch-style github profile cards pip install termfetch 149/mo

Lore.

• · Made my first internet money winning an online hackathon Google ran — prize money, and a box of goodies.
• · Finished Freedom Fighters in six hours using every cheat code the game had, then got banned from its servers for exploiting it online.
• · Played chess up to district level, and won a stack of tournaments getting there.
• · Taught chess to a lot of younger players while still getting beaten by better ones myself.
• · Bluffed my way into a cybersecurity internship on experience I didn't have; spent the first weeks quietly catching up, and finished it clean.

Socials.

GitHub npm PyPI LinkedIn Email