$30K
Follow-on contract won
A proof-of-concept became production web + Android apps — and a signed $30K extension.
97% RECALL@5300MS QUERIES5,000+ PHOTOS INDEXED
FPT SOFTWARE · INTERN · 2024
NVIDIAINCOMING SWE
AdobeAMBASSADOR
RunpodGRAND PRIZE
NSFREU RESEARCH
IEEEPRESENTED @ MIT URTC
WPIBS + MS '27
UC Berkeley2ND · AI HACKATHON
Stanford3RD · AI HACKATHON
NASASPACE APPS · PEOPLE'S CHOICE
FPTAPPLIED AI · $30K
Zolli LabsFOUNDER
CHAPTER 01 — RECEIPTS
Proof, with sources.
Every number here traces to a contract, a prize, a release, or a paper — nothing decorative.
$8K
Runpod Grand Prize
FlashML took the grand prize at Runpod's hackathon — then kept shipping instead of stopping at the demo.
GRAND PRIZE — TOP OF FIELD$8K CASHSHIPPED AFTER THE WIN
RUNPOD HACKATHON · 2026
100+
Pilot users on PyPI
FlashML ships as installable Python packages with full CI/CD — real users, not a repo trophy.
2,200+ TESTSPYTORCH · SKLEARN · HF ADAPTERSGITHUB ACTIONS CI/CD
ZOLLI LABS · FLASHML · LIVE
−60%
Forecast error, with RAG
RAG-grounded LLM feature selection over 157 exogenous features — presented at IEEE MIT URTC.
157 FEATURES−27% VS XGBOOST BASELINE400+ REPRODUCIBLE RUNS
NSF RESEARCH · IEEE · 2025
CHAPTER 02 — THE WORK
Four real projects.
Open one.
Only shipped, verifiable work makes the shelf. Select a project and OPEN unfolds the full case study.
GRAND PRIZE · RUNPOD
2ND · UC BERKELEY AI HACKATHON
EDGE AI · 163MS
01 / 04FlashMLSelf-hostable distributed ML across heterogeneous compute — 47% faster batches, survives worker failures.ARROWS · SELECT · OPEN
CHAPTER 03 — THE ROAD
Five stops. One line each.
2024
FPT Software
Applied AI intern — multimodal image search for 5,000+ photos.
97% RECALL@5 · $30K CONTRACT
2025
NSF Research
REU — LLM-guided feature selection for data-scarce forecasting.
LLM-TSFS · IEEE MIT URTC
2026
Adobe
AI product tester on Acrobat Student Spaces, invited back as ambassador.
TESTER → AMBASSADOR
2026 · NOW
Zolli Labs
Founder — building FlashML in the open, from prize to product.
FLASHML · 100+ USERS
NEXT
NVIDIA
Software engineering — AI-assisted GPU validation workflows.
INCOMING SWE
HONORS — $10K+ PRIZE VALUERUNPOD GRAND PRIZESTANFORD AI HACK — 3RDUC BERKELEY AI HACK — 2NDNASA SPACE APPS — PEOPLE'S CHOICEGOATHACKS ×2GPA 3.92 · PRESIDENTIAL SCHOLARSHIPWPI BS CS + MS AI · '27
CHAPTER 04 — NOW
Making mismatched machines act like one.
FlashML's pull-based scheduler lets fast workers claim more jobs instead of assigning work blindly — batches finish 47% sooner across a 3.7× speed range. When a worker dies, its jobs go back to the queue instead of being lost.
Next chapter: AI-assisted GPU validation at NVIDIA.
THE ACTUAL FLASHML MODEL — NOT A DECORATIVE GRAPHLIVE
JOB QUEUE — 128 PENDING
MACBOOK · SLOWCLAIMED 17 JOBS
GPU TOWER · FASTCLAIMED 61 JOBS
GPU CARD · DIED28 JOBS → BACK TO QUEUE
BATCH COMPLETION VS STATIC SCHEDULING47% faster