Aleksi

(or Alex, if you wish)

Building thingsTbilisi

Intro

I build things, mostly the parts that aren’t visible from outside. A few projects at the moment, all of which started because the existing thing was wrong in a way I could name precisely.

Before that: competitive programming, freelance work, and a route into the industry I had to construct myself, since there wasn’t one where I grew up.

Work

belay

belay records a filesystem’s state, diffs two records, and explains the difference. Packages come from pip’s and dpkg’s own databases, branch moves and history rewrites from git’s refs and reflog, text changes as line counts. Permission and capability changes are never folded away.

Given a task’s protect spec, it checks the end state against a baseline taken before the run. It sees state, not what a process read. A scan of a live tree marks what changed under it as unknown instead of guessing.

Dogfooded on its own development. Open source.

C11Linuxsingle static binarypre-alpha

Two records, one diffTwo small file listings on the left, labelled before and after, recorded. Lines from both converge on a card on the right, explained, holding three lines: pip, numpy 2.1.3 to 2.2.0; git, main rewritten by rebase; and, highlighted, a question mark: build/ unknown, changed during scan.beforeafterrecordedpipnumpy2.1.32.2.0gitmain rewritten by rebase?build/unknown: changed during scanexplained
Fig. 1Two records, one semantic diff

DoesThisPass

DoesThisPass judges strength lifts from video: squat, bench, deadlift, pull-up and dip. A pose model tracks body landmarks in each frame, reps are segmented from that motion, and every rep is checked against fixed rules: hip crease below the knee, elbows locked out, chin above the bar line. Each rule has a pass threshold, a fail threshold and an unsure band between them.

Unsure calls go to a human, as does anything the video can’t show: an occluded limb, a bad camera angle, the load on the bar, the lifter’s bodyweight. It does not read plates. Thresholds are starting values, not yet tuned on labelled footage.

The judge and a review queue run locally. Public leaderboards and a phone app are planned.

Pythonpose estimationffmpegSupabasepre-alpha

The callA stick figure at the bottom of a squat, side view. A short bracket measures the gap between the hip crease and the top of the knee, and a line drops from it to an axis labelled hip crease versus top of knee, in centimetres. Two thresholds split the axis into fail on the left, a hatched unsure band in the middle that goes to a human, and pass on the right. The measured rep lands as an amber dot in the unsure band.failpasship crease vs top of knee (cm)unsurehuman
Fig. 2The call

Cabin

YouTube’s problem isn’t the recommender. It’s what the recommender is ranking. Run the same algorithm over a corpus that was gated on the way in and the output changes completely.

Cabin is that gate. Every submission is watched end to end by a person before it’s let in - not sampled, not skimmed. The bar has four clauses: it promises something, it delivers exactly that, everything it attempts is good, and it’s worth the time of whoever it’s for.

Currently a test run. Embed-only, measuring one thing: whether curated discovery is a habit people repeat.

live

The gateSixty grey points on the left stand for everything submitted. Each points toward a narrow aperture between two bars; only five pass through it and arrive as amber points on the right.most infew through
Fig. 3The gate

Background

CompetitiveCodeforces Expertpeak 1719
ICPC NERC regional finalstwice
Google Kickstarttop 300
EducationWarsaw University of Technologycoursework, no degree
WorkHuaweiML research internship
Clorofounding engineer
GoogleSRE, passed the interview loop

Elsewhere

© 2026 Aleksi Svanidze