Ishaan Ranjan

Efficiency research on NVIDIA’s Nemotron-Elastic models.1 I measure how accuracy trades against latency across their elastic subnetworks: which of the smaller models inside the large one are worth running, and what each one costs in quality.

Luxen’s products have reached more than 5,000 users. I built and shipped two open-source projects there. Arbor, formerly AgentTree, is a desktop workspace for agent trees, shared context and live multi-agent coding sessions.2 It gained more than 1,250 users after its January 2026 release. Regimen is an orchestration layer for small language models: it breaks a long task into parts and routes each to a specialized small model.

We’ve raised more than $48,000 in grants and cloud credits, including a 1517 Fund Medici grant.

The usual way to cut the cost of a multi-agent system is to swap its expensive worker model for a cheap general one, and it quietly fails. We fine-tuned a 3B model for one narrow role and built a harness around that role. On 200 held-out code-navigation tasks per condition, drawn from real repositories, it reached 94.5% success against 95.5% for a 72B baseline, at 84.8% lower cost per successful task.3 The white paper is [1].

A voice-first iPhone app for reporting construction defects. A worker describes the problem out loud; the phone transcribes it with Whisper and structures it with Gemma 3n, both on the device, ties it to the building’s BIM model, and syncs it through Supabase for the architect to review.

A 51-jurisdiction panel (50 states and DC) of alcohol policy from 2000 to 2022, an event-study difference-in-differences design4 to estimate what each policy change did, and a public dashboard for the results. The paper is on SSRN [2]. The project was the basis of a $5,000 BASIS Educational Ventures scholarship.

I analyzed how vitamin D, omega-3 and exercise relate to epigenetic aging clocks, end to end in R and Python. The question I presented: do vitamin D supplements go with slower biological aging? Across 428 people and three clocks, I found no significant association.5 Sex, BMI and education were associated with the aging measures; supplementation was not. I presented it in a poster talk to more than 800 people [5].