Vladimir Osipov
NAME
vladimir osipov - co-founder of Excession Labs, mathematics student at UCL
vladimir@osipov.cc · github.com/alygg77 · linkedin.com/in/vladimir-osipov123
A random process draws my name in dots.
DESCRIPTION summary
- NOW
- I co-founded Excession Labs. Our AI agents test the security of a client’s systems.
- BEFORE
- I co-founded Project25. We signed a US$350,000 financing agreement.
- WON
- Our team of three won the AI Engine hackathon, among more than 200 participants.
- STUDY
- Mathematics and Statistical Science at UCL, to June 2027.
EXPERIENCE
Excession Labs
co-founder · October 2026 to nowA security lab. Our AI agents attack a client’s systems, with permission, and find the weak points. excessionlabs.ai
Penetration tests, security audits, AI red teaming.
- We also make AI models run 4 to 10 times faster on NVIDIA and AMD GPUs.
LLM inference. AMD Instinct MI350X/MI355X.
Project25 Inc.
co-founder · May 2025 to October 2026An autonomous system that made, published and measured content on real devices. project25.live
- I incorporated the company. We signed a US$350,000 financing agreement with HomeDAO in December 2025.
A SAFET.
- I ran the AI image generation on rented GPUs and fixed its start-up faults.
ComfyUI on Runpod. Cold starts, CUDA.
- From September 2026 we worked with YourEra (US) on adverts made by users, and measured which adverts brought sales.
ReflectAI
co-founder · February to March 2025Three of us built an AI agent that spoke with a company’s customers by phone. It won the AI Engine hackathon.
- After the win we had investor conversations with Bessemer Venture Partners and Zero Prime.
P10Y
productivity research · January to July 2025I did customer research for P10Y, which studies the productivity of software engineers: lists of target companies, and scorecards for local businesses.
- I worked with Yegor Denisov-Blanch (Stanford Software Engineering Productivity study), under a P10Y NDA.
AWARDS
AI Engine: UK University Hackathon
London · February 2025Overall first place, and first place in workplace automation, with team ReflectAI. More than 200 participants from UCL, Oxford, Cambridge and Imperial.
Prize: NVIDIA H200 GPU access.
ETH Oxford.
Third-place prize from Ava Labs, US$800, with team NLWallet. February.
AI for Good hackathon.
Second place. Eurasian Hub, November.
School olympiads, Russia.
Prize winner, National Data Analysis Olympiad (DANO). Certificate of honour, National Mathematics Olympiad.
Other hackathons.
2024 to 2025: EasyA × Stellar, AI Builders (Eurasian Hub), Imperial and UCL Hack, EF Hack.
PROJECTS my own, since June 2023
NewtonCards.
An AI flashcard platform, 2024 to 2025, built in the UCL AI Foundry incubator (Cohort 3).
StellaRamp.
Moves money from African mobile wallets to the Stellar blockchain. Built at the EasyA × Stellar hackathon.
Rust, Svelte.
A small model of it. Clockwise, it stops and turns back. Flick it, or:
Smaller projects.
A campus-tours app on the App Store (code), a Python app that finds leads on the web, and a script that trades on the news.
EDUCATION
University College London
London, UK · expected June 2027BSc Mathematics and Statistical Science, third year.
Moscow State School 57
Moscow, Russia · June 2024Mathematics track with accelerated studies. GPA 4.8 of 5.
- In 2023 I was a senior volunteer teacher there. I taught evening mathematics classes, ran the annual geometry test and wrote two inter-school olympiad problems.
IMS School
Limassol, Cyprus · June 2024International A Levels: A*A*A. I taught myself Mathematics, Further Mathematics and Economics.
- MAT, the mathematics admission test of the University of Oxford: 78 of 100 (exam only).
RESEARCH drafts, not published
Five working papers for AAAI-27 workshops. I am writing three more.
AI judges give higher scores to longer forecasts, not to more accurate ones.
Longer, Not Better · PDF · no workshop yet
When an AI model plays a person, its answers change at the wrong times.
Simulated Users Get the Time Course Wrong · CALE-AI
A forecast shows what comes with an action, not what the action causes.
When Is a Known Future Covariate an Intervention? · PDF · AI4TS
An AI model trained before the Iran war reads the President’s posts on it. We test its reading against the oil price 10 minutes later.
Reading the President’s Iran-War Posts with a Pre-War-Cutoff LLM · PDF · FinFM
AI bank-support agents follow their written procedure even when a step in it is unlawful.
The Procedure Comes First · PDF · FinFM
NOTES short articles
I write them to check that I understand a thing.
But do LLMs think?
8 November 2024I rephrased standard coding tasks and tested GPT-4 again: does it solve them, or remember them?
Elliptic Curves in 5 minutes
6 October 2024How adding points on a curve gives you keys and signatures.
zk-SNARK: QAP
5 August 2023How an equation becomes a proof that hides its inputs.
ENVIRONMENT skills and interests
| LANGUAGES | English (fluent, TOEFL 107), Russian (native). |
|---|---|
| PROGRAMMING | Python, C++, Swift, LaTeX. Beginner: Golang, R, Rust, machine learning. |
| TOOLS | OpenAI, ElevenLabs and Twilio APIs, Next.js, Flask, Selenium, ComfyUI, serverless GPU (Runpod). |
| INTERESTS | Debating (UCL Debating Society), sports, crypto and blockchain, travelling, politics, hackathons, venture capital. |
CONTACT
| vladimir@osipov.cc | |
| GITHUB | github.com/alygg77 |
| linkedin.com/in/vladimir-osipov123 |
BUGS errors on this page
Probably some, in the code of this page. If you find one, please tell me.
The rest is for engineers.
SEE ALSO this page in a terminal
less(1), mandoc(1), man(7).
This page is also a real manual file, vladimir.7:curl -sO osipov.cc/vladimir.7 && man ./vladimir.7
COLOPHON how this page works
| FILE | One HTML file of 75 kB, and three font files. No framework, no tracker, no cookie. |
|---|---|
| NAME IN DOTS | A Metropolis-Hastings sampler with 40 chains. |
| SPINNING TOP | A rigid ellipsoid that rolls and does not slip (Kane and Levinson, 1982), with friction. Fourth-order Runge-Kutta, 1,200 steps a second. |
| KEYS | As in the pager less. Press h for the list. |