Neelay RanjanGenerative modeling for domains where labels are scarce and mistakes are expensive.

NASA Ames Research Center · Regenstrief Institute · Purdue University

I build generative models for domains where labeled data is scarce and a wrong prediction costs more than a missed one. My first-author paper on label-efficient vessel segmentation from catheter angiograms is in preparation. At NASA Ames I work on airspace coordination and a synthetic dataset for air traffic control speech; at Regenstrief I trained the diffusion model behind that paper. I also built a 553 KB chess engine; you can play it further down this page. I'm on a gap semester from my AI degree at Purdue, applying to master's programs for fall 2027.

Table 1. Headline numbers from the paper and the chess engine.

0.882Dice at 16 labeled angiograms, from my paper in preparation

25/25paired runs ahead of every baseline at 16 labels

~75%faster surgeon corrections

553 KBchess engine, roughly 1900-2200 Elo

Research

In my first-author paper, in preparation with Shantanu Dev and Andrew Gonzalez at Regenstrief Institute, I trained a diffusion model to predict the clean segmentation mask directly instead of the noise. It reaches 0.882 Dice from 16 labels, ahead of five baselines in 25 of 25 paired runs. Trained the usual way, predicting noise, the same architecture stays near 0.23 Dice at every budget.

The same model made a surgeon ~75% faster at correcting its output.

x axis is log-spaced. budgets 16 (5%), 32 (10%) and 80 (25%) labels of the 320 training images.

Figure 1. Mean test Dice at each label budget, pooled over every seed and fold (2,500 predictions per point). At 16 labels x0-diffusion leads the next best model, zero-shot SAM, by 0.047; by 32 labels DeepLabV3 and ResNet-UNet have caught up. ε-diffusion (~0.23) is left off the axes.

Pelvic-iliac angiogram from the segmentation benchmark, test image 196 — **Figure 2.** Every per-image test Dice at 16 labels, pooled over all seeds and folds (2,500 per line). At 0.5 Dice, 0.32% of x0-diffusion’s predictions fall below, against 5.4% for SAM and 13.5% for ResNet-UNet. x0-diffusion’s mean is only a few points ahead of the baselines, but it almost never fails outright. The images come from a seed-1 re-run, one per slider stop, chosen by SAM Dice near the cursor; SAM is stochastic, so its mask can score differently from its recorded run.

**Figure 2.** Every per-image test Dice at 16 labels, pooled over all seeds and folds (2,500 per line). At 0.5 Dice, 0.32% of x0-diffusion’s predictions fall below, against 5.4% for SAM and 13.5% for ResNet-UNet. x0-diffusion’s mean is only a few points ahead of the baselines, but it almost never fails outright. The images come from a seed-1 re-run, one per slider stop, chosen by SAM Dice near the cursor; SAM is stochastic, so its mask can score differently from its recorded run.

I also led the PCB design team for a helical antenna for electromagnetic field stimulation in Alzheimer’s disease therapy; we presented it as an oral at IEEE MWSCAS 2026 in Cincinnati on August 11.

NASA Ames Research Center

On SLAAC, space launch and air/airspace coordination, I trained a diffusion model on FAA radar tracks to reroute flights around closed airspace. It learns what routes look like offline and applies the airspace while it samples, so a new closure needs no retraining. A reroute has to come back as an ordered list of named fixes a controller would accept, and the operator chooses how far ahead it may look; the figure below has both settings.

Since August I’ve been on SHIFT: speech-to-text for air traffic control that makes fewer errors than the Whisper system in use, and a generate-and-rank parser that turns the transcripts into a maneuver database. I’m writing the parser up with NASA as an extended abstract for AIAA AVIATION.

dataPublic where I could, at some loss of quality: the airspace is the FAA’s Special Use Airspace layer, plus past launch TFRs for the sites with nothing charted, since NASA’s airspace file isn’t public. With no historical route database, the filed routes are my flight-plan model’s.

differencesMy SLAAC models are integrated into the simulation software NASA and the FAA use to evaluate future air traffic management strategies. This demo keeps the diffusion model’s weights but samples in fewer steps than my pipeline does, on public airspace and routes written ahead of time. A reroute takes seconds on my laptop, longer with more airspace or every flight at once. On this demo’s test cases, fewer final reroutes keep every leg the full buffer distance from the airspace than the poster reports, though almost none cross it.

Live systems

Both demos below run the trained models in your browser.

Experience

since Aug 2026NASA Ames · SHIFTSpeech-to-text for air traffic control, a generate-and-rank parser from transcripts to maneuvers, and a synthetic speech corpus that grows the recorded data about 20x.active

May-Aug 2026NASA Ames · SLAACDiffusion rerouting around closed airspace, and a flight-plan language model that writes a day of about 44,000 flights.complete

2024 to presentRegenstrief Institutex0-diffusion vessel segmentation and a synthetic angiogram pipeline, with Shantanu Dev and Dr. Andrew Gonzalez.active

expected May 2027Purdue UniversityB.S. in Artificial Intelligence, Intelligent Control & Systems concentration, math minor. John Martinson Honors College, Dean’s List, 3.68 GPA.gap semester

2026Alzheimer’s stimulation antennaLed the PCB design team on a helical antenna for electromagnetic field stimulation; presented as an oral at IEEE MWSCAS 2026.complete

2025Davinci WearablesAn agentic vision pipeline that estimates nutrition from meal photos, under 15% error.complete

2024-2025V2X aircraft-maintenance LLMLed the two-stage RAG design; cut hallucinations from about 40% to about 5%.complete

Figure 6. Green is active now; amber is paused until January.

References

Bootstrapping surgeon labeling campaigns with x0-diffusion (in preparation)
N. Ranjan, S. Clarke, et al., “Generate-and-rank extraction of air traffic control clearances into an en route ontology,” extended abstract for AIAA AVIATION 2027 (in preparation)
F. Perez, J. Morisaki, H. Kanakri, M. Rizkalla, et al. (incl. N. Ranjan), “Helical Antenna for Electromagnetic Field Stimulation in Alzheimer’s Disease Therapy,” IEEE MWSCAS 2026 (oral)
Resume
GitHub
ORCID
LinkedIn
Email
Supplementary material

The background is the sky over NASA Ames.

The sky over NASA Ames from the moment you opened the page. Objects are drawn enlarged, in long-exposure color.