Empirical ML researcher and engineer. SWE @ Google Ads, ex-NASA, Goldman Sachs. Bi-coaster (SF & NJ). 10x Hackathon Winner. ECE (Electrical & Computer Engineering) + Math at Rutgers-New Brunswick. I work on safe, interpretable AI: LLM evaluation infrastructure, quantum-inspired learning, and multimodal medical imaging.
Currently at Google (Ads AI/ML), NASA (swarm robotics for lunar infrastructure), and Columbia Center for AI (quantum ML research). Founder of the Grey Matter Society at Yale School of Medicine — 150+ chapters worldwide.
$5.5M raised across projects with NEC, Nokia, Google, Qualcomm, Robert Wood Johnson Hospital.
- Google Ads · YouTube AI/ML — Simulation-based forecasting in Python and C++ across 20K+ advertising entities, on distributed infrastructure scaling toward 50M+ simulated user-query scenarios.
- Swarm Robotics Intelligence Project · NASA-funded, Rutgers — Decentralized, fault-tolerant multi-agent software for a heterogeneous robotic swarm; autonomous coordination and self-assembly for lunar infrastructure. Delivering technical reviews directly to NASA stakeholders.
- Columbia Center for AI · IEEE · INNS — First-author empirical ML research benchmarking quantum-inspired and classical optimization methods on noisy medical-imaging data.
- Founder & President · Grey Matter Society — Yale School of Medicine. 150+ chapters globally.
- NASA SpaceTech · L'Space Academy — Led systems architecture for InnerSolace, an AI circadian-rhythm modeling platform simulating 100+ mission-day scenarios across orbital and polar environments. 1st nationally.
- Goldman Sachs · SWE Emerging Leader — Selected first-year cohort. Backend prototyping for low-latency, high-throughput data pipelines; performance and reliability tradeoffs in production-grade systems.
- WINLAB (Wireless Information Network Lab) — Real-time simulation and telemetry infrastructure for XR systems. Asynchronous data pipelines, structured logging, distributed tracing.
- National Science Foundation · Princeton — Cross-platform backend translating insights from 40+ user interviews into measurable reliability improvements for financial systems.
cross-sae — Interpretability: do vision transformers and the human visual cortex share the same features? Trains sparse autoencoders on both a ViT and human brain responses to the same images, then matches features across domains under false-discovery-rate control (Model-X knockoffs). The cross-domain feature match is an honest null at current SAE capacity — with a diagnostic (RSA ρ = 0.155, p = 0.0005) showing the shared structure is real and recoverable. Error-controlled, self-correcting empirical interpretability.
Quantum ML for AML detection · arXiv:2601.18710 · code First-author publication, January 2026. Equilibrium Propagation reached 86.4% accuracy without backpropagation; a 4-qubit Variational Quantum Circuit held its accuracy with 5× less training data. Established reproducible QML baselines on the 18,365-image AML-Cytomorphology dataset. Rutgers + Columbia.
Aurelis — Reproducible LLM-as-judge evaluation infrastructure: it grades clinical notes against a rubric and validates the AI grader against human reference scores (quadratic-weighted kappa, Pearson) on real ACI-Bench clinical notes — applied scalable oversight, with a content-addressed cache for byte-identical reruns.
Thaakat (EndoDetect) — Multimodal radiomics decision-support for endometriosis: an EfficientNet + XGBoost ensemble on pelvic MRI with Grad-CAM explanations, behind a strict non-diagnostic boundary.
Circa — InnerSolace, productized: a chronotype-aware scheduler built on a tested two-process alertness model that places demanding work at your biological peak (recovers up to +203% modelled throughput for evening types vs. clock-blind scheduling).
Robrick — The NASA swarm self-assembly work as runnable, tested code: identical units build and self-heal a target structure from one neighbour-only local rule (decentralized gradient assembly, Kilobot-style).
Fitra · repo Circadian-optimization app for Muslims. The five prayers are treated as immovable anchors; everything else — sleep, nap, caffeine cutoff, peak focus — is optimized around them. Two-process sleep model, live recompute, present mode.
AeroBin + routing engine — Multi-sensor, edge-compute smart-waste system; the prediction + optimization engine cuts wasteful pickups from 87% → 0.5% and servicing events −70%. 1st Place, national Verizon Smart Campus Competition (NEC · Nokia · Google · Qualcomm).
- TransformerLens #1369 — a Direct Logit Attribution tool for mechanistic interpretability.
- SAELens #697 — covariance-whitening normalization for training sparse autoencoders.
- Inspect Evals #1765 — registering the MedCalc-Bench medical-reasoning benchmark in the UK AI Safety Institute's evaluation framework.
- 1st nationally — NASA SpaceTech Competition
- 1st Place national — Verizon Smart Campus Competition (AeroBin)
- Rutgers Innovation Award · Outstanding Student Leader (20 of 60K+)
- Elected Representative to 40K+ students — Rutgers School of Engineering · RUSA
