Skip to content
View scharfff's full-sized avatar
:shipit:
Focusing
:shipit:
Focusing

Organizations

@SintezDev

Block or report scharfff

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
scharfff/README.md

Telegram LinkedIn Google Scholar ORCID Web of Science

ML engineer. NLP and RAG systems for technical and industrial domains.


English · Русский

I build LLM-based document analysis systems: RAG over regulatory documents, OCR pipelines, structured extraction with validation, fine-tuned BERT-like models for NER and classification. Most of this work was done on closed corporate documents and included comparing local models against cloud providers.

My first degree is in civil engineering (BIM), so I know technical documents from the inside. I also teach Python, computer science and mathematics to university applicants at SPbSTU and do NLP research.

Projects

 

DocuLex: contract analysis service (closed source)

  • Problem. Reviewing a contract by hand took about 15 minutes.
  • Architecture. PDF, DOCX and scans go through an OCR pipeline. Structured extraction returns 17 fields validated against a Pydantic schema, and the service produces PDF and XLSX reports. It has a Telegram bot interface and is deployed in Docker on Amvera.
  • Outcome. I built it at Sintez on my own initiative. Review takes about 5 minutes instead of about 15, and the service is still in use at the company.

RAG over technical regulations at Sintez (closed source)

  • Problem. Search over technical regulations from Russian electrical codes and standards (PUE, PTEEP, GOST, RD) in a closed corporate corpus of about 300 documents.
  • Stack. LangChain, ChromaDB, ruBERT embeddings, a FastAPI service, Langfuse for tracing, Docker.

electrolabbot: open RAG assistant for electrical testing standards

  • Problem. A working RAG system for electrical testing built only on public standards and synthetic reports, so no private data is published.
  • Architecture. The corpus is public standards plus 15 anonymized synthetic reports generated by a script. ruBERT embeddings and a Chroma store with an HNSW index run locally, and the assistant keeps conversational memory. The LLM is called through the Hugging Face Inference API. Langfuse tracing is optional. Interfaces: Streamlit, a FastAPI REST API and a CLI. Docker and Amvera configs are included.
  • Trade-off. It uses a free-tier hosted LLM. The README documents a limit of 30,000 requests per month and 2 to 5 seconds of latency, and lists caching and asynchronous handling as mitigations.

TapKeys: trackpad gestures to shortcuts on macOS (Swift, MIT)

  • Architecture. Touches come from Apple's private MultitouchSupport.framework, loaded at runtime with dlsym. A CGEvent tap swallows a click only if it uses three or four fingers and has a binding. Letter shortcuts follow the current keyboard layout, including AZERTY, Dvorak and Russian.
  • Constraints. Private APIs can change between macOS releases. The app is tested on macOS 15 (Intel), signed ad hoc and not notarized. The README says so and gives workarounds.
  • Release pipeline. CI runs on every pull request and every push to main. A version tag builds a universal binary, a zip and a DMG. The build checks the signature and both architectures and runs a self-test from the DMG. The pipeline creates the release as a draft, verifies checksums of the uploaded files against SHA256SUMS, then publishes it and updates the Homebrew cask. v0.2.0 is the first stable release: brew install --cask scharfff/tap/tapkeys.

spbpu-inf-course: web textbook for the 2027 Russian Unified State Exam (EGE) in computer science (content in Russian)

  • Architecture. An Astro web app built and deployed by GitHub Actions to GitHub Pages. Content lives in per-task folders with Python solutions. The previous year is kept on its own branch (2026), and the README documents the draft 2027 specification: 27 tasks and 29 primary points.
  • Outcome. The core of my educational platform and the material for my classes at SPbSTU. Open the web app.

mailrucloud: unofficial CLI for Mail.ru Cloud (PyPI)

  • Architecture. Python and Click over WebDAV. In v1.1.0 I replaced the outdated REST API with WebDAV. Multithreaded upload (4 threads by default) arrived in v1.4.2, and the CHANGELOG reports that sync is 4 to 8 times faster.
  • Engineering. Tests run with pytest, CI covers Python 3.10 and 3.11, and releases go to PyPI through a publish workflow: pip install mailru-cloud-cli.

Research

Master's research on linguistic markers of personality type in texts: MBTI-LSTM, a BiLSTM classifier in PyTorch. I also contributed research to a CNN-based bird recognition system for surveillance camera snapshots.

Experience

 

  • 2025 to present. Lecturer (part-time), SPbSTU, applicant outreach center. My own course and its web platform.
  • 2024 to 2025. ML Engineer, Sintez, IT Infrastructure team. RAG system, DocuLex, fine-tuned BERT-like models for NER and classification of test reports, a fine-tuned Hugging Face model that summarizes electrical laboratory documentation.
  • 2023 to 2024. Analyst-Developer, Sintez, IT Infrastructure team. Benchmarks of local models against cloud providers (OpenAI, Hugging Face, Sber, Yandex), A/B tests of ML and AI solutions with follow-up analysis.
  • 2022 to 2024. Research Engineer (part-time), SibSIU, Department of Information Technology. LSTM and BiLSTM classifier research.
  • 2022 to 2023. Analyst intern, Sintez. ETL processes, dashboards, digitizing technical documentation per GOST.
Publications

 

Publications are in Russian and appear under the name Samoilova L. (surname now Scharf). The name variants are linked in Google Scholar. Web of Science ResearcherID: NZO-1576-2025.

  1. Samoilova L., Gulevich T. M. Application of an artificial neural network with the Big Five personality model to personnel selection. Proc. All-Russian Conf. "Science and Youth: Problems, Searches, Solutions", iss. 28, part II. SibSIU, 2024.
  2. Samoilova L., Danilovsky V. M. Prediction of Personal Characteristics MBTI Using Recurrent Neural Network LSTM and Text Data of Social Networks. Vestnik nauki, 2024, vol. 4, no. 6 (75), pp. 1176 to 1184. Full text. Code: MBTI-LSTM.
  3. Samoilova L., Danilovsky V. M. Intelligent Bird Recognition System in Video Surveillance Based on Convolutional Neural Networks. Vestnik nauki, 2024, vol. 2, no. 11 (80), pp. 1149 to 1159. Full text.
Education

 

  • MSc in Information Systems and Technologies (Intelligent Systems and Technologies), SibSIU, 2022 to 2024.
  • BSc in Industrial and Civil Engineering (BIM), SibSIU, 2018 to 2022.
  • English: B2.
Stack

 

  • ML: Python, PyTorch, Hugging Face Transformers, fine-tuning (LoRA, PEFT), vLLM, Ollama
  • LLM and RAG: LangChain, LlamaIndex, Chroma, Qdrant, Langfuse, PaddleOCR, Pydantic
  • Services: FastAPI, Docker, PostgreSQL, Linux, Git, GitHub Actions, Yandex Cloud, Cloud.ru, Amvera
  • Apps: Swift (TapKeys)

Pinned Loading

  1. mailrucloud mailrucloud Public

    Unofficial CLI for Mail.ru Cloud over WebDAV: upload, download, directory sync. On PyPI as mailru-cloud-cli

    Python 7 1

  2. electrolabbot electrolabbot Public

    Open RAG assistant for electrical standards (PUE, PTEEP, GOST). ruBERT, Chroma, Langfuse, FastAPI, Streamlit (docs in Russian)

    Python 4

  3. MBTI-LSTM MBTI-LSTM Public

    BiLSTM classifier of MBTI personality features from social network text. PyTorch, master's research

    Jupyter Notebook 2

  4. spbpu-inf-course spbpu-inf-course Public

    Handbook for Russian USE in Computer Science (Scharf L. Course for SPbSTU). Task walkthroughs, Python solutions, progress tracker

    Python 1

  5. tapkeys tapkeys Public

    Free macOS menu bar app that binds trackpad taps and clicks to shortcuts. Swift, Homebrew cask, MIT

    Swift