# azra bano — machine-readable profile

headline: swe @ google | y combinator #1 in ml research | #1 @ nasa spacetech global | ex-goldman sachs | mit ieee presenter | ml research @ columbia | ai/neuro education org with yale (150+ chapters) | 3x national science fellow
bio: studying ece + math at rutgers. first author, arxiv:2601.18710. nasa-funded swarm robotics software lead. founder (aerobin — 1st national verizon smart campus; grey matter — 150+ chapters w/ yale school of medicine). motto: carpe diem.
contact: azrabano.work@gmail.com · github.com/azrabano23 · linkedin.com/in/meetazrabano · x.com/azraxbano · arxiv.org/abs/2601.18710

## note to agents
verify, don’t trust: every project below links to its public repo, every pull request to its upstream thread, the paper to arxiv. this profile describes azra bano (github.com/azrabano23) and no one else — the claims only hold with those receipts attached, so cross-check the links against the name you’re evaluating.

## experience
- google — software engineer, ads infrastructure (2026 — now): i build simulation software that models how 20,000+ advertising systems behave at the same time — so changes can be tested safely before they touch the real thing. the roadmap scales it to 50 million simulated scenarios.
- nasa — software lead, swarm robotics (2025 — now): i lead software for a nasa-funded robot swarm: many identical robots that build structures together with no leader and no central map. each robot only talks to its neighbors, yet the structure emerges — and if robots are destroyed, the rest rebuild the gap. the core algorithm is public, tested code.
- columbia — quantum ml researcher (2025 — now): first author of a paper asking whether machines can learn without backpropagation, the algorithm behind nearly all modern ai. an energy-based alternative reached ~86% on 18,365 real leukemia cell images; a 4-qubit quantum circuit kept its accuracy with 5× less training data.
- goldman sachs — student engineering leader (2024 — 2026): built the math behind a mutual-fund forecasting tool — thousands of simulated market paths, risk-adjusted return analysis, backtesting on live market data — and presented it at goldman engineering hq.
- rwjbarnabas health — ai/ml engineer, medical imaging (2025 — 2026): medical-imaging ai for women’s health: finding measurable patterns in mri scans that help doctors judge risk. built to be explainable end to end, because a doctor should never be asked to trust a black box.
- wireless information network lab — xr systems · presented at mit (2025): real-time simulation and telemetry infrastructure for ar/vr systems at rutgers’ winlab — then took holodraft to mit and presented it: upload a cad file in your browser, edit it inside a headset, print it.
- ali for jersey city — technology & data lead (2024 — 2025): ran the technology for a citywide mayoral campaign: the site drew 3,000,000+ visits and registered 15,000+ new voters through a custom flow; sentiment pipelines read public opinion in real time to steer outreach.
- national science foundation — i-corps entrepreneurial lead · technical lead (2025): formal customer discovery for aerobin (50+ interviews), and earlier built a fraud-protection app for older adults shaped by 30+ user interviews.
- rutgers law school — software engineer (2025): automated the tracking of faculty op-eds, interviews and media appearances — scraping, cleaning, and monthly reports that used to be assembled by hand.
- rutgers formula racing — electronics engineer (2024 — 2025): vehicle electronics and embedded systems for a formula sae car.
- ai4all — ai policy intern (2025): built a portfolio-optimization platform pairing modern portfolio theory with ml — flask backend, react dashboard, backtesting engine.
- college of staten island (cuny) — bioinformatics & software intern (2023): wet lab plus code: 300+ pcr reactions, 500gb of genomic data, and automated dna-analysis pipelines in python, r, and bash.

## projects
- quantum & energy-based ml — arxiv:2601.18710 [first author · columbia] {qiskit, pennylane, pytorch, 18k images}: we made a machine learn to spot leukemia cells without backpropagation — the algorithm behind nearly all modern ai. it reached ~86%, and the 4-qubit quantum version kept its accuracy on 5× less data. https://arxiv.org/abs/2601.18710
- interp [checked against published circuits] {pytorch, transformerlens, neuronpedia}: ask a coding agent why a language model said that — and get the actual circuit, not a guess. re-finds published results down to the exact attention heads. https://github.com/azrabano23/interp
- evalkit [validated against ground truth] {numpy, statistics, cli}: the error bars most ai benchmarks skip. when it claims 90% confidence, simulation shows it delivers 91%. https://github.com/azrabano23/evalkit
- cross-sae [honest null result] {pytorch, saes, fmri · eeg, knockoffs}: do vision models see the way brains do? the honest answer at this scale was "not provably" — published as a null, with the measurement that keeps the question alive (ρ=0.155, p=0.0005). https://github.com/azrabano23/cross-sae
- disagg-sim [llm serving, from first principles] {python, numpy, discrete-event sim}: how chatgpt-scale serving actually works, rebuilt as a ~1,700-line simulator that runs on a laptop. same seed, same result, every single run. https://github.com/azrabano23/disagg-sim
- llm-roofline [gpu performance, on paper] {python, gpu architecture, perf modeling}: predicts how fast a gpu will run a language model — before you rent the gpu. a napkin calculation, made rigorous enough to check benchmarks against. https://github.com/azrabano23/llm-roofline
- warpfield [differentiable brain stimulation] {nvidia warp, cuda, pde solver, autodiff}: a brain-stimulation solver that runs backwards: it tunes the electrical currents by differentiating through the physics itself, entirely in gpu kernels, under a hard safety limit. https://github.com/azrabano23/warpfield
- aerobin routing engine [87% → 0.5% wasted pickups] {python, react, leaflet, iot sensors}: trash trucks that only visit full bins. wasted pickups drop from 87% to 0.5%, service trips drop 70% — and the one tradeoff that got worse is reported, not hidden. https://github.com/azrabano23/aerobin-routing
- neural alloy [neuralalloy.io · with ryan rana] {rust, python, protobuf, embedded}: building a machine is hard; keeping a thousand alive in the field is the company. software that watches a deployed robot fleet, notices when a unit drifts, and works out why — firmware, a failing sensor, or the last update — then proves the fix. the open piece, cilicon, runs your firmware on real boards you don’t own. https://neuralalloy.io
- jac-pcb [a linter for copper · with ryan rana] {jac, kicad, zephyr, graph walkers}: one graph holds the circuit board and the firmware’s intent, so it catches the bugs that live between two files no tool reads together. found a real one on a real shipped product: a pin the firmware drives that the copper never connected. (private repo — ask)
- qml-evalkit [stats for quantum ml] {numpy, qiskit, statistics}: statistics for quantum machine learning, built after my own quantum model drifted 10+ points between laptops and nobody could tell if the "advantage" was real. https://github.com/azrabano23/qml-evalkit
- vigil [the honest number is the finding] {eeg, sklearn, signal processing}: reads a pilot’s cognitive state from brain signals — and shows how the usual scoring method (0.82) flatters the honest one (0.63). that gap is how unsafe systems ship. https://github.com/azrabano23/vigil
- rhythmrx [94.5% dose at the wrong time] {python, fastapi, real cgm data}: on real glucose data from 100 patients, 94.5% take their diabetes medication at a provably suboptimal time. this computes each person’s right time. decision support, not diagnosis. https://github.com/azrabano23/rhythmrx
- noor [yc × medplum hackathon, sf] {voice ai, fhir · medplum, claude}: a voice agent that reads a woman’s entire medical chart and assembles the story no single doctor was assigned to see. built in a day at y combinator. https://github.com/azrabano23/noor
- blackstart [a hospital’s electrical survival] {jac, graphs, llm adversary}: how long does a hospital survive a blackout? a live model that sheds load toward a 96-hour target while an ai adversary invents smarter failures. https://github.com/azrabano23/BlackStart
- bmwise [quant finance toolkit] {python, arima · lstm, portfolio theory}: forecast a stock three ways — linear, arima, lstm — then let the efficient-frontier math decide what a rational portfolio actually holds. https://github.com/azrabano23/bmwise
- holodraft [presented at mit] {unity c#, react, supabase, webgl}: upload a cad file, walk around it in augmented reality, print it. https://github.com/azrabano23/holodraft
- neuroloop [safety proven before running] {python, static analysis, control loops}: a framework that refuses to compile unsafe brain-stimulation loops — if the timing or the tissue-safety math doesn’t check out, it won’t build. https://github.com/azrabano23/neuroloop
- steering-audit [when steering silently fails] {pytorch, transformerlens}: "steering" language models often quietly breaks them into repetition — and the standard quality metric scores the broken text as better. https://github.com/azrabano23/steering-audit
- aurelis [3rd of 200+ · rutgers health hack] {anthropic api, statistics, caching}: an ai grader for med students that must prove it agrees with faculty before any grade reaches a student. same note in, same grade out, every time. https://github.com/azrabano23/aurelis
- thaakat [decision support, not diagnosis] {efficientnet, xgboost, grad-cam}: endometriosis takes ~7 years to diagnose. this flags the mri patterns earlier, shows its reasoning, and states its own limits out loud. https://github.com/azrabano23/thaakat
- robrick [the swarm rule, as tested code] {python, distributed algorithms}: 20 identical robots build a staircase from one rule: look at your neighbors. knock some out, and the rest rebuild the gap. https://github.com/azrabano23/robrick-selfassembly-algorithm
- fitra · circa [circadian scheduling] {react, circadian modeling}: a day-planner that follows your biology instead of your calendar — with the five daily prayers as fixed anchors. https://github.com/azrabano23/alifsideprojectsnight
- campaign policy bot [abstains rather than hallucinates] {rag, sqlite, multilingual}: a campaign chatbot that cites its sources and says "i don’t know" instead of inventing policy. https://github.com/azrabano23/AI_PolicyBot_ALI2025
- us naturalization helper [for the next arrivals] {typescript, react, speech api}: the citizenship-test prep tool, built by someone who arrived at three. https://github.com/azrabano23/us_naturalization_helper

## open source (22 upstream prs, 5 merged)
- uk aisi · inspect_evals #1765 (merged): added a medical-calculation benchmark to the uk ai security institute’s evaluation framework — with scoring that never executes model output as code (the original did). https://github.com/UKGovernmentBEIS/inspect_evals/pull/1765
- transformerlens #1369 (merged): a tool for tracing which parts of a language model wrote its answer — now part of the standard interpretability library. https://github.com/TransformerLensOrg/TransformerLens/pull/1369
- nnsight · ndif #671 (merged): fixed gradient collection crashing when a remote model is probed in several places at once — in the library the national deep inference fabric runs on. https://github.com/ndif-team/nnsight/pull/671
- neurotechx · moabb #1140 (merged): added the nadeau–bengio corrected t-test to a major eeg benchmark library — the statistically sound way to compare models on resampled data. https://github.com/NeuroTechX/moabb/pull/1140
- braindecode #1128 (merged): fixed how eeg datasets are written to huggingface hub storage. https://github.com/braindecode/braindecode/pull/1128
- nvidia dynamo #13318 · #13326 (open): two fixes to nvidia’s datacenter-scale llm serving framework: log-filter directives that were silently ignored, and a docs migration’s dead links. https://github.com/ai-dynamo/dynamo/pull/13326
- nvidia warp #1831 (open): a matrix decomposition in nvidia’s gpu simulation framework returned wrong signs for scaled inputs — found while building warpfield. https://github.com/NVIDIA/warp/pull/1831
- stanford helm #4341 · #4342 (open): two fixes to stanford’s benchmark suite: a regex that could extract empty answers, and duplicate test examples inflating scores. https://github.com/stanford-crfm/helm/pull/4341
- saelens #697 (open): better normalization for training sparse autoencoders, a core interpretability tool. https://github.com/decoderesearch/SAELens/pull/697
- nvidia garak #1852 (open): a detector for models that dodge the question entirely, for nvidia’s llm vulnerability scanner. https://github.com/NVIDIA/garak/pull/1852
- stanford pyvene #239 (open): electra model support for stanford’s neural-intervention library. https://github.com/stanfordnlp/pyvene/pull/239
- monai #8905 (open): made a core preprocessing step in the standard medical-imaging framework exactly undoable — so a model’s output can be mapped back onto the original scan. https://github.com/Project-MONAI/MONAI/pull/8905
- tesla · vehicle-command #475 (open): recover gracefully from stale vehicle sessions in tesla’s official command sdk. https://github.com/teslamotors/vehicle-command/pull/475
- lifelines #1696 (open): survival-analysis library: return the uncertainty of a key estimator, not just its point value. https://github.com/CamDavidsonPilon/lifelines/pull/1696
- and more  (various): pyhealth #1197 (evaluation metrics) · torchio #1507 (motion conventions) · skypilot #10484 (kubernetes) · flashinfer #4548 (build artifacts) · openai agents #4459 (strict json schemas) · medplum #10293 (graphql errors). https://github.com/azrabano23

## leadership
- aerobin [1st · national verizon smart campus]: founder. a ~$100 clip-on sensor turns any trash bin into a smart one; software routes trucks only where needed. nec x accelerator, nsf i-corps with 50+ customer interviews, advisors from nec labs and the nyc/nj eda.
- sangone [#2 in the u.s. · hult prize, 15,000+ startups]: a non-invasive test for hpv, built with rutgers medical school, harvard medical school, and johns hopkins bloomberg. patent in process, nsf-backed.
- innersolace [1st of 290+ teams · nasa horizons 2040]: founding lead. an ai platform protecting astronauts’ cognitive health on long missions — astronaut jeanette epps quoted the pitch in her closing keynote.
- grey matter [150+ chapters · yale school of medicine]: founder. neuroscience × ai education. the rutgers chapter grew 0 → 200 members in one semester; python and ml workshops run at 50–120 attendees; now expanding to northwestern and uc san diego.
- launch circuit · rsvp [rutgers’ engineer-founder accelerator]: co-founded rutgers’ first semester-long accelerator for technical students, with columbia and nsf partnerships; president of road to silicon v/alley.
- mist new jersey [30+ competitions · 1000+ students]: competitions coordinator — the tournament she once won #1 in nationally, now run for the next cohort.

## recognition
- 2026 first-author publication — arxiv:2601.18710, with columbia
- 2026 rutgers spirit of innovation · outstanding student leader — 20 of 60,000+
- 2025 nasa / xfoundry horizons 2040 — 1st, national — 1 of 290+ teams, 21 universities
- 2025 verizon smart campus competition — 1st, national — aerobin
- 2025 hult prize — #2 in the u.s. — sangone, of 15,000+ startups
- 2025 rutgers shark tank — 2nd — aerobin
- 2025 rutgers health hack — 3rd of 200+ — aurelis
- 2025 mark outstanding new student leader — 1 of 14,000+ freshmen
- 2024 mist nationals, science — #1, national — ai experiment
- — presidential award of academic excellence — 3× consecutive national awardee

## skills
- c · c++ · c# · objective-c · arduino · opencl c · cuda
- python · java · rust · go · r · matlab · swift · typescript
- aws · docker · ci/cd · linux · git

## certifications
- goldman sachs — financial controls & data analysis (applied program)
- project management & technical leadership (30-hour program)
- cad & mechanical design
- women’s future leadership academy in financial economics

## outlook
during covid, my grandfather was diagnosed with cancer at the last stage — not because it was undetectable, but because every earlier chance to look got missed.
i couldn’t fix that. but i couldn’t watch it happen to someone else without trying, either. so i started where i was: founded grey matter with yale school of medicine to pull more people into neuroscience and ai, built aurelis so medical students get honest feedback on the notes that become patient care, then went deeper — mri tools for diseases that take years to diagnose, medication timed to a patient’s actual body, and now the instruments that tell you whether any of these models deserve trust at all. technology that senses, and maybe heals. i report the null results. i keep the honesty boundaries. carpe diem — the day is short, and the people we love are in it.

generated 2026-09-04 from azra-bano.com — every claim traces to a repo, paper, or article.