83% of data recovered against greedy's 69% · 0.00% gap on every instance the exact solver closed

satellites make more data than they can get down, so something has to choose which passes to use. i built the whole chain: sgp4 propagation from real tles, coordinate transforms, visibility windows, link budgets, then a simulated annealing scheduler over the passes that fall out. the part i care about is that i also wrote a branch and bound solver, so on the instances small enough to close i can prove the schedule optimal instead of saying it looked good.

false shutdowns 130 to 6 · faults found 30% to 70% · 904 flights

an engine controller shuts an engine down when a measured channel leaves its band. it has flown for sixty years and it cannot tell a failing engine from a failing pressure transducer, because both look the same from inside the band. i wrote the ascent simulator, the fault library, five detectors built on different assumptions, and the harness that scores all of them over identical flights.

268.7M real messages · whole session in 95 MB under a collector that never reclaims

a limit order book and matching engine in java that never allocates on the hot path. it reads raw nasdaq itch, rebuilds the book, and matches against it. real data falsified my core design assumption: almost all of a session fits in 21,654 ticks, and then a tail of sub penny stink bids demands twenty million. no synthetic workload would have shown me that.

65,536 fp8 pairs checked against exact arithmetic

i rebuilt the arithmetic of a matrix multiply in java, bit for bit, so every decision inside it becomes its own setting: number format, scaling, rounding, accumulator width, even the order the products get added in. on a gpu you cannot change one and hold the rest still. the point is attribution, so when a low precision run dies you can say which decision killed it.

labeled eval harness · top 3 retrieval

full stack semantic search over your screenshots, the graveyard of information nobody can search. the interesting part isn't the search, it's the labeled eval harness i built to grade top 3 retrieval accuracy and find where it was quietly failing.

5,000+ monthly users · 12 countries

an ai + cybersecurity learning platform: browser based linux terminals on docker, ctf challenges, and courses. i handle the isolation, resource limits, and teardown so thousands of strangers can get a shell without breaking anything that matters.

6 attack families · 8 defenses · hermetic sandbox

how often prompt injection actually works on an agent that can use tools, and which defenses move the number. six attack families against eight defense variants, scoring attack success and task completion together, because a defense that blocks everything also blocks the work.