• Role: Builder AI & Data Systems
• Education: BITS Pilani, Goa → University at Buffalo (2+2) · B.E. EEE
• Building: Closr
• Focus: Agents · Orchestration · Inference cost · Workflows
• Mantra: Build things that save you money, compute, and time.
• Creative problem solving under resource constraints

Production-grade, OpenAI-compatible AI gateway engineered to load-balance LLM requests across free-tier providers (Gemini, Groq, Cerebras). Built as an inference resilience layer featuring Redis-backed distributed circuit breakers, dynamic model virtualization, token accounting, and SHA-256 request caching to maximize $0 inference. Architected in a pnpm monorepo using Turborepo and published directly to the public npm registry as an open-source infrastructure package.

Fully autonomous B2B lead generation pipeline built for the creator economy, engineered to run entirely on local hardware (4GB VRAM RTX 3050). Scrapes 7 high-intent web sources and processes the DOM through a 4-stage Extraction Fortress (using CPU-bound bart-large-mnli and all-MiniLM-L6-v2) to aggressively filter noise before a single token touches the GPU. Extracts entities via qwen2.5:7b (Ollama) and routes them through a concurrent 5-thread ReAct Agent state-machine for waterfall email enrichment—maintaining strict $0 cloud LLM costs.

Identity verification pipeline for the creator economy that proves platform ownership without requiring credentials. Engineered a 4-layer verification architecture — OAuth root-node proof, secondary-link ingestion, queue-based scraping, and local LLM structured analysis — to produce tamper-resistant creator profiles backed by crawl-verified metrics rather than self-reported data. Built as a full-stack monorepo with a Next.js frontend, Supabase database contracts, public profile pages, and a background worker pipeline for async scraping and analysis queues.

Palantir-grade intelligence pipeline that scrapes and structures geopolitical data into a Neo4j knowledge graph. Engineered a 2-Layer Resolution Gauntlet using fuzzy matching and all-MiniLM-L6-v2 vector cosine-similarity to disambiguate entities. The entire inference engine is memory-optimized to run within a strict 4GB VRAM ceiling via dynamic context reduction and a rolling keep-alive window.

Secure event ticketing ecosystem that eliminates fraud using TOTP (Time-based One-Time Password) entry mechanisms with 30-second refresh cycles. Implements resale intelligence and smart-contract royalties to track and verify all legitimate ticket transfers. Live on Vercel.

Languages:

AI / Data:

Infra / Deployment:

"The best systems are the ones you never notice — because they just work."