alteraltoml / ai · portfolio
Sheet 01 Title sheet AT · REV 2026

ML / AI Engineering

AlexeyTkachenko

ML/AI Engineer — LLM · RAG · Production

I build production-oriented LLM and RAG systems across cloud and on-prem environments.

Strongest at evaluation, reliable failure handling, observability, and the engineering around the model.

TG @altarch MAIL xelartn75@gmail.com GIT github.com/alteralto
Sheet 02 Build log · P-01 — P-05 AT · REV 2026

01 / Build log

Selected projects

Systems built from research and evaluation through infrastructure and delivery.

P-01 · Flagship

ArchNorm

Founder · Full-cycle engineer · Feb 2026—present

A production B2B RAG system for automated compliance review of construction documentation, deployed in Yandex Cloud and as an offline desktop edition.

  • Hybrid retrieval over a 20-document regulatory corpus with mandatory clause attribution and citations.
  • Quality evaluation uses a 132-question benchmark, golden sets, retrieval metrics, and refusal controls.
  • Production hardening includes provider failover, observability, monitoring, backups, and offline inference.
REGULATORY CORPUS 21 PDF · SP / SNIP PROJECT DOCUMENTATION PDF HYBRID RETRIEVAL BM25 + DENSE CLAUSE ATTRIBUTION CITATIONS REQUIRED REPORT EVIDENCE-FIRST NO EVIDENCE — REFUSAL DELIVERY: CLOUD · OFFLINE DESKTOP
Fig. 01 — Compliance review loop
Specification
B2B
RAG product
21
source PDFs
2
delivery modes
EVAL
evidence-first
PythonFastAPIRAGHybrid retrievalDockerYandex CloudTauri

P-02 · Embodied AI

Melissa

ITMO Persona Lab · 2025—2026

Voice-and-agent R&D system combining STT → LLM/RAG → TTS, local inference, game commands and world state, plus Twitch, Discord, OBS, and VTube Studio integrations.

AgentsRAGLocal inferenceIntegrations

P-03 · Tabular ML

MIPT Digital Department hackathon

Team captain · Mar—Apr 2026

Led a four-person team under NDA. Built a three-branch ensemble for out-of-time evaluation, reaching churn recall 0.98 (166/170), 1st place of 8, and an expert score of 9.7/10.

PythonEnsemble MLOut-of-timeDashboard

P-04 · DevTools

SourceCraft

ITMO × Central University · 2026

AI skills that detect and semi-automatically resolve YAML/JSON conflicts across pull requests, with CI integration, log triage, reports, branch updates, and notifications.

AI skillsCIYAML / JSONGit workflows

P-05 · Product analytics

VK Case Champ · S-Commerce

Solo case

EDA and category segmentation for conversion and GMV, followed by a placement strategy with a projected GMV uplift of 25–30%.

EDASegmentationConversionGMV
Sheet 03 Capabilities · S-01 — S-06 AT · REV 2026

02 / Capabilities

Engineering toolkit

Model quality matters. The surrounding system makes it usable.

S-01

LLM / RAG

RAG architectures · BGE-M3 · GigaChat · YandexGPT · ChromaDB · Qdrant · FAISS · BM25 · hybrid retrieval · query rewriting · VLM table extraction

S-02

Evaluation & safety

Hit@K · MRR · NDCG · golden sets · refusal metrics · confidence calibration · guardrails · provider failover · observability

S-03

Agentic AI

Tool calling · LangGraph · smolagents · MCP · skills · domain-agent orchestration · conflict resolution

S-04

Backend / Data

Python · FastAPI · Pydantic · pytest · SQL · pandas · NumPy · scikit-learn

S-05

Infra / Desktop

Docker · Compose · Yandex Cloud · Linux · Caddy · TLS · GitHub Actions · backups · uptime monitoring · llama.cpp · Ollama · LM Studio · GGUF · Tauri

S-06

Languages

Russian — native · English — B2

Sheet 04 Education AT · REV 2026

03 / Education

Education

Architecture domain knowledge reinforced by formal machine learning training.

  1. 2021—2026

    Kuban State University

    Architecture · GIS/BIM · AI/ML applications in architecture and urbanism

  2. 2025—2026

    MIPT · Foundations of ML/DL

    Professional program · 270 hours · final grade 5/5

  3. Additional

    Applied programs

    MIPT Phystech.Idea · ITMO MegaSchool / AI Talent Hub · BootcampLabs SQL · Cloud.ru · Process Mining