Ankit Bari
AI Engineer | Multimodal and Multilingual model | Model Architectures, Mathematical Foundations & Custom CUDA Kernels
- •35% Researching, 30% Building, 35% Inference & Deploying
- •Trying to build models (Multimodal and Multilingual) at the best possible with the resources I have!
- •Not done yet, optimize GPU for training and model for production
- •Currently Building: SkyBoy
EExperiences
4 mos
Duration
Jun 2026
Launched
- •A fast, searchable, agent-agnostic directory of portable AI skills (SKILL.md) and plugins.
- •Ships as a website, docs site, dual-distribution CLI (npm + PyPI), and an MCP server so any skill is one command away from Claude Code, Cursor, Windsurf, Gemini CLI, Codex, ChatGPT, and Claude Desktop.
6 mos
Duration
Dec 25 - May 26
Period
- •Founder of SiteVPN, a privacy SaaS combining private browsing, disposable email, anonymous temporary phone numbers, AI-assisted privacy intelligence, subscriptions, and a Business developer API in one account.
- •Built as a Next.js App Router product with Prisma/PostgreSQL, NextAuth, Stripe subscriptions, Claude-powered AI features, and API-key protected REST endpoints for VPN, temp email, and temp number access.
6 mos
Duration
On-site
Mode
- •Developed an AI chatbot for lead tracking, achieving 95% accurate user data capture.
- •Implemented a fintech authentication backend with low-latency (150-200 ms) request handling for 100+ daily users.
PProjects
SkyBoy
Agent-agnostic directory of AI skills (SKILL.md) — preview, install, and ship them to Claude, ChatGPT, Cursor, and Gemini CLI.
Fermes
Sarvam-LLM land-suitability agent that fetches live geospatial data via tool-calling, grounded in verified USGS/FEMA/USDA data to cut hallucinations.
Mini-Interfaze: Receipt Field Extractor
CRNN encoder + Transformer decoder for receipt field extraction (vendor, date, total, address) with bounding boxes. Custom fused CUDA kernel (RMSNorm + residual) speeds up training.
Wispr Flow Clone
Two-stage pipeline: Whisper for verbatim transcription, then an LLM cleans fillers, grammar, and tone across email/chat/notes. Custom fused CUDA kernel (bias-add + GELU) speeds up inference.