Screenshot model distillation
A 7B-to-2B vision-model experiment, with a reproducible training and evaluation pipeline. A small proof of concept, including where the metrics misled me.
We could work together.
An AI feature to evaluate, a first product to build, or a system that needs to serve thousands of people. I help teams turn those problems into software they can test and use.
We start with a useful outcome and a way to check it. Then we scope a first version, work through the architecture and implementation, and use the results to decide what comes next. I make the technical trade-offs clear in terms of the product, the team and the people using it.
Tell me what you’re building, where you’re stuck and what you need to change.
sergey.paarfenov@gmail.comA bit about my work
I joined Yandex Praktikum as the 11th member of its early product team. I helped build the product, designed the test architecture for a coding trainer used by 7,000–10,000 students a day, and developed an ML model to assess project submissions before ChatGPT.
At TripleTen, that work expanded into AI assessment, search and tutoring. The problems were concrete: shorten grading time, help students find the right material and make practice easier to produce. I worked from backend architecture through production releases, contributed to learning personalisation and the Mastery Depth Tracker, and helped launch the QA Engineering programme in the US.
As Co-Founder and CPO of IAWY, I took on product direction as well as development. We built an AI tutor with automated project review that integrated into existing learning platforms, and launched it in two bootcamps in Latin America.
Today, I’m applying that mix of product and engineering at Aliwio, where I’m Co-Founder and CTO, building a CRM for independent veterinary clinics. As Tech Adviser at Symptomato, I help shape the technology for a service where AI gathers context before a human health specialist joins the conversation.
In the open
Agent memory & provenance
An experiment harness for measuring how memory compaction changes an agent’s decisions — including results that contradicted my own hypothesis.
Where I can help
- An AI idea worth testing
- Choose a narrow problem, define a baseline and build a working prototype. Use the results to decide whether AI improves the product enough to justify further development.
- Building an early product
- Build the first version across architecture, backend and interface. Work with the team to choose what belongs in it, what can wait and what needs to hold up when people start using it.
- Agents you can evaluate
- Define correct behavior, inspect failures in tool use, memory and retrieval, and build evaluations the team can repeat. Check recovery as well as the successful path.
- Smaller models, measured trade-offs
- Test distillation and fine-tuning for a specific task. Compare quality, latency and memory use against a baseline on the same workload, then decide whether a smaller model earns its place.
- Learning platforms & automated assessment
- Build test architecture for coding trainers, ML assessment of project submissions and useful feedback. Connect those systems to the curriculum, where learners get stuck and what practice they need next.
Education & courses
Stanford University
Saint Petersburg State University of Finance and Economics
Computer Information Systems
The path here
Aliwio
Co-Founder and CTO
Feb 2026 to present
Responsible for technical architecture and product engineering, building a CRM for independent veterinary clinics.
Symptomato
Tech Adviser
Jan 2026 to present
Advise on technology for a service that combines AI intake with human support: an AI assistant gathers context before a health specialist joins the conversation.
TripleTen
AI Engineer
Dec 2021 to Jan 2026
Built and shipped AI systems for assessment, search and tutoring, connecting backend engineering with the learning experience.
Designed and implemented a plan and execute assessment system that replaced 140 legacy text-area quizzes. Grading latency fell from 8.4 to 0.6 seconds, a 14× speedup.
Built a CMS vectorization pipeline on Kafka processing 100,000 items a day, and delivered semantic search end to end. Search engagement increased by 9%; time to content fell by 18%.
Automated practice generation, cutting task authoring from 2.5 hours to 18 minutes. Built an internal VS Code plugin that reduced average support time per student issue by 27%.
Worked on Voice Mode for AI Tutor (Dot), integrating WebRTC, speech recognition, an LLM and speech synthesis. Helped bring it to a stable release with 97.8% successful sessions and a median of 1.3 seconds to the first response.
Contributed to learning personalisation and built features that improved student retention. On the Mastery Depth Tracker, I connected product and engineering work to measure learners’ confidence and identify where they needed more practice. Also helped launch the QA Engineering programme in the US.
IAWY
Co-Founder and CPO
Aug 2023 to Jul 2024
Co-founded IAWY and developed an AI tutor with automated review of student project submissions. Defined the product direction and roadmap for a tutor that integrated into existing learning platforms. We launched it in two bootcamps in Latin America.
Retailhub
AI Researcher
Oct 2023 to Jan 2024
Designed a pipeline to enrich product listings: normalize source data, extract attributes, generate structured features and produce descriptions in natural language.
Yandex Praktikum
Software Engineer
Mar 2018 to Dec 2021
Joined as the 11th member of the product team and helped build its early components, including the systems used to test code and assess student work.
Designed the test architecture for the coding trainer used by 7,000–10,000 students a day.
Developed an ML model to assess student project submissions before ChatGPT, and contributed to the launch of the Web Developer programme.
Earlier experience
Thingyfy
Software Engineer
Oct 2019 to Oct 2020
Built startup products with React and TypeScript.
Yandex School of Data Analysis
Learning Specialist
Oct 2018 to Jan 2020
Worked as a learning specialist in Moscow.
Sravni.ru
Software Engineer
Nov 2017 to Oct 2018
Built insurance products in React.
Earlier work
Engineer and teacher
2016 to 2019
European Gymnasium
Computer science and web development.
Spaces
Commercial property product engineering.
MCS
Arduino, Swift, Python, web, and VR teaching.