Updates from Aditya Karad

Fall 2026

I built Costco Index to track prices for everyday essentials across U.S. Costco warehouses. You can compare regions on a map, look back at price changes, and keep a watchlist. Costco's low margins and high volume made me curious whether those prices could reveal small shifts in consumer costs.

I'm also interested in whether the data could help with investment research on companies exposed to changing ingredient costs, like Starbucks and Chipotle. I wouldn't treat it as a national inflation measure. Promotions, package sizes, membership requirements, and local availability can all affect the prices I see.

I attended NorCalDB Day at UC Santa Cruz.

Instinct Relay. A local MCP tool that lets coding agents use Instinct's generous daily token budget for small coding tasks over WhatsApp. It returns diffs for review. The source code is on GitHub.

I forked OpenAI's Codex CLI and integrated Jev to build /autoroute, an LLM router that picks a model and reasoning effort for each task from a model catalog. It aims to cut model costs and also suggests skills. I ran evals for both; the skill selector found the expected skill in 76 of 84 prompts. I put my fork on GitHub and wrote up how it works and what I tested.

I pitted Jev against GPT 5.6 Luna (medium) in a heads-up Texas Hold'em poker game and found that Luna mogged Jev.

On average, Jev took a decision in about 250ms, whereas Codex took ~7s. Jev's style was cautious and reactive. It rarely raised on good hands and often folded with the best hand. I found that people on my Twitter timeline were using Jev in places where it doesn't belong, so I thought it would be interesting to test its ability in long horizon games against a reasoning model like Luna.

Aditya launched Podcast Index to compare the companies represented by guests on the Acquired and Cheeky Pint podcasts.

Summer 2026

Aditya explored coding-agent harnesses, user interfaces, and agent architectures. At Oracle, he built and supported a browser-first coding-agent harness whose browser UI worked with the CLI installed on the machine. It brought together conversations, prompts, skills, memories, MCP setups, and AGENTS.md instructions. He kept it working with the internal models available to his team, added phone controls and completion notifications, and let coworkers join a shared agent thread through an exposed local port. He also built keyboard shortcuts, macros, and internal tool integrations. He rebuilt EarlyApply around agents that generate and repair job-board scrapers, maintain generated scraper code, and run parallel workers. EarlyApply and a Notion-native version reached the finals of the Notion Hackathon. He also began more structured cardio training and studied the Johns Hopkins poker course.

Summer 2025

Aditya joined Oracle's Database group as a Member of Technical Staff on the GoldenGate team.

2023–2025

Aditya completed an MS in Computer Science at UC Irvine, worked on Apache Texera and database coursework, and served as a teaching assistant. In summer 2024 he worked at AppLovin on backend systems. He previously worked full-time at JPMorgan Chase from July 2022 through August 2023.

For the complete chronology, use the HTML updates page.