Перейти к содержимому
Портфолио

Кейсы и проекты

10 AI-кейсов с интерактивными демо, архитектурными диаграммами и метриками.

Projects

AI-инженерия и стартапы

От автономных фабрик приложений и agent-платформ до гибридного RAG и multi-model consensus. Каждый проект — доказательство инженерного паттерна.

Most technically complex
🤖AI / Solo R&D

mapp_conveyer — Autonomous App Factory

Built solo: personal R&D to validate autonomous AI systems at production depth. Nine specialist agents handle ideation → research → spec → code → review → compliance → publish. A board of 4 reasoning models (Claude Opus 4, GPT o4-mini, Gemini 2.5 Pro, DeepSeek-R1) deliberates on every release decision. DAG-scheduled parallel codegen with kill-gates placed before expensive stages. 27/27 automated quality gates passed on the first run — each gate a specific, verifiable assertion about pipeline correctness.

9 agents · 4-LLM board · 27/27 quality gates
PythonFastAPIClaude Opus 4GPT o4-miniGemini 2.5 ProRedis StreamsPostgreSQL
Agent infrastructure
🕸️AI / Solo R&D

AI Agent Platform

Built solo: the infrastructure layer under the other agentic builds. LangGraph for stateful multi-step agent graphs, LiteLLM as a unified gateway across providers, and Google's A2A protocol for agent-to-agent interop. Gateway + UI with per-agent routing, tracing and cost control.

LangGraph · LiteLLM · A2A protocol
PythonLangGraphLiteLLMA2AFastAPI
🧠AI / Solo R&D

RAG Knowledge Graph

Built solo: a private knowledge base that answers natural-language questions over proprietary documents with exact source citations. Three retrieval methods (Qdrant vector, BM25, Neo4j graph) fused via Reciprocal Rank Fusion enable multi-hop reasoning beyond vector similarity. Evaluated with RAGAS — faithfulness 0.67 → 0.91 on 200 control questions.

Hybrid retrieval + RRF · RAGAS 0.67→0.91
PythonLangChainNeo4jQdrantBM25RAGAS
⚖️AI / Solo R&D

Council — Deliberation Service

Built solo: an on-prem deliberation service in Go (stdlib only, closed-box). Multiple LLM 'members' reason in parallel on the same question, then their positions are reconciled — disagreement is surfaced, not averaged away. Built for environments where data cannot leave the perimeter.

Parallel LLM members · on-prem · stdlib-only
GoLLMOn-premConcurrency
🗞️AI / Solo R&D

Real News — Anti-Doomscroll Aggregator

Built solo: a news aggregator that clusters ~20 sources with HDBSCAN, then rewrites each story with an LLM and runs an automated bias-check pass before publishing on a fixed daily schedule. Unsupervised clustering plus evaluation — not an infinite feed.

HDBSCAN clustering · LLM rewrite + bias-check
PythonHDBSCANDeepSeekTelegram
📡AI / Solo R&D

AI Radar — Daily AI Digest

Built solo: an automated daily AI-news radar. Collects from Hacker News, HuggingFace, GitHub, Reddit and RSS, deduplicates, re-ranks with an LLM, and ships a curated digest to a Telegram channel every morning — orchestrated in n8n with full tracing via Langfuse.

5 sources → dedup → LLM rerank → daily digest
n8nDeepSeekLangfuseTelegram
🎧AI / Solo R&D

NZT — Real-Time Dialogue Navigator

Built solo: an AI real-time dialogue assistant for negotiations and interviews. Whisper transcribes speech live, GPT-4o generates contextual hints, WebSocket streams them to an Android app, and an FSM adapts prompts to the dialogue phase. 200+ tests across backend and mobile.

Real-time STT + GPT-4o · 200+ tests
PythonFastAPIKotlinWhisperGPT-4oWebSocket
🛡️AI / Solo R&D

Autonomous Security Recon Agent

Built solo: an autonomous pipeline for authorized security testing. Recon → scanning → out-of-band verification → triage, with an LLM prioritizing findings and a two-layer authorization gate enforcing scope and rate limits before any request. Includes an IDOR-detection engine and a Next.js dashboard. Runs only against explicitly authorized, in-scope targets.

Recon → OOB-verify → triage · scope-gated
PythonDeepSeekNext.jsSecurity
🖼️AI / Solo R&D

Multimodal Listing Pipeline

Built solo: a multimodal pipeline that turns a photo into a ready marketplace listing. A vision model recognizes the item, an LLM drafts bilingual listing cards (title, description, price), a pricing step calibrates against live comparables, and a human approves before publish. Human-in-the-loop by design; 110 tests.

Vision → bilingual cards → pricing → HITL
PythonGemini VisionDeepSeekHITL
⛏️Startup

Job Miner

Built solo: an AI job-search assistant that turns hours of manual browsing into a 2-minute morning digest. Scrapes and deduplicates 10+ job boards daily, scores each position against a profile via a semantic ML pipeline (3-LLM pseudo-labeling + ablation study), and pushes a curated top-10 to Telegram.

10+ job boards · semantic scoring · daily digest
PythonFastAPIPostgreSQLTelegram Bot

Готовы обсудить?

Если вам нужен архитектор, который строит автономные AI-системы — напишите.

Сербия (Белград) · CET/CEST · рабочие часы совпадают с EU · Опыт международных контрактов