I'm Varun Jakkula — an AI engineer and founder. I ship LLM-native products end-to-end, from the model layer to the landing page, and I get them into the hands of real users. Founding engineer at a 2M-user product, creator of a platform that hit 20,000 users in 45 days.
Building a stealth, AI-powered learning platform — an AI "teacher twin" for K-12 students, aimed at thousands of schools and districts across the United States.
Led AI strategy and product transformation for a legacy presentation-template company serving 15,000+ businesses and 350,000+ users — bringing an AI-first product roadmap to an established brand.
Engineered high-performance LLM response parsers cutting generation time to under 30s via streaming; built an in-browser editor powering 600,000+ presentations (3,000+ added weekly); shipped two AI image-gen apps on Flux LoRA in a one-month window; optimized a 2M-row Postgres database (77% faster planning).
Built high-converting landing pages, custom dashboards, intake forms, AI chatbots and end-to-end lead funnels for home-services, moving and logistics companies across the US, UK and Canada — with CRM, analytics and automation wired in.
A test-series platform for engineering-entrance exams. Scraped 22,700+ questions with Playwright into an RLS-secured Postgres, then drove growth through creator-led UGC campaigns — going from a standing start to a real business in six weeks.
A production-grade AI scribing tool that auto-generates medical charts in seconds using ElevenLabs and custom transcription models — saving physicians 20+ hours a week. Grown to 100+ signups across 4 countries via targeted Facebook ads.
Editing an LLM's memory with fine-tuning — durably writing new knowledge into a model so it recalls facts across rephrasings while keeping the rest intact. Step one toward stateful-memory LLMs.
A diffusion-based generative model built on a custom U-Net with hand-written noise samplers, trained from scratch to synthesize pixel-art sprites.
A CLI coding agent that writes code and scaffolds whole projects from natural language — built in Feb 2024, a full year before Claude Code shipped.
A complete end-to-end RLHF alignment pipeline implementing the full spectrum of modern preference-learning algorithms — SFT through DPO and PPO.
An LLM evaluation, annotation and benchmarking platform for measuring and comparing model behavior systematically.
A Chrome extension that saves temporary chats from ChatGPT, Claude, Gemini, Grok and 10+ AI tools — so you never lose a conversation that turned useful.
An AI fact-verification system using hybrid, modified RAG over video transcription and web search. Indian Patent Office — App. no. 202441071795.
Authored and published "Evaluating Machine Learning and Deep Learning Models for Accurate Stock Price Prediction."
Contributed CI workflow configs verifying compilation, formatting and Clippy linting for the open-source encryption service.
I'm open to AI engineering and founding roles, and to interesting collaborations. The fastest way to reach me is email.