AI that earns its keep.
AI agents, smart search (RAG), and automation — live in production, not slideware. Try it.
Opens Echofy — my live SaaS, running in production.
The proof, live.
Echofy
An AI employee for Instagram shops.
A production multi-tenant SaaS I designed and shipped solo. It reads a shop's Instagram DMs and answers product questions, qualifies buyers, and handles support 24/7 — in the customer's own language, across text, images and voice.
Projected It's built to take the ~80% of DMs that are the same repeat product questions off a shop owner's plate — they only step in for the rest.
It's live and running in production right now. Don't take my word for it — try it yourself.
Open the live demo ↗Proven property: a shop can never be left without an owner — checked by symbolic execution (CrossHair, Z3-backed) over all code paths, not sampled cases.
In plain terms: it reads photos and voice notes, replies in each customer's own language, and keeps every shop's data walled off from the rest.
I use AI to build AI.
Echofy is a full production SaaS I shipped solo. That's only possible because I turn the same agentic techniques I build for clients inward — onto my own engineering. It's how I ship solo, at the pace of a team. This is the working method behind the build.
My own agent process
Custom skills and subagents I built to run production development end to end — solo.
custom skills · subagentsWired into real systems
MCP servers connect my agents straight to the tools, data and APIs a real build touches — not a sandbox.
MCP serversOrchestration at the frontier
I work with Claude Code's newest layer — Dynamic Workflows: plan a job, fan it across parallel subagents, self-verify, merge.
Dynamic WorkflowsPushing the harness further
Between client builds I push my own agent harness — longer-lived sessions, durable cross-session memory, recursive LLM architectures. Experiments that feed straight back into how I ship.
context · memory · recursive LLMsPractical AI, built properly.
Scope is agreed up front, with a first version live in 1–2 weeks. We start small and paid, and keep going only if it earns its place.
AI Knowledge Assistant
Finds and delivers answers across a team's documents and systems. Covers data and process analysis, a custom RAG architecture, a web or messenger interface, and team onboarding.
See related work →AI Sales Chatbot
A 24/7 sales agent that answers product questions, qualifies leads, books meetings, logs to a CRM, and hands off to a human when needed.
Optional: Stripe payments · custom integrations
See related work →Custom AI Workflows
Automates repetitive work like document sorting, email and form processing, and reporting. Covers process analysis, building the workflow, integrations with existing tools, and documentation/handoff.
See related work →Built to prove it.
Builds I've taken end to end to push the stack and pressure-test ideas. Every one actually runs.
Workflow Autopilot
An autonomous Telegram bot that sorts ~10,000 telecom equipment photos a month with a two-stage AI pipeline — 1.5 hours of manual work down to a 2-minute review.
Automated Sales
A conversational sales agent that answers course questions from a RAG knowledge base, qualifies leads in real time, and closes with in-chat Stripe checkout — the whole funnel in one chat.
Smart Ticketing
An autonomous LangGraph agent that classifies tickets, answers from a RAG knowledge base behind a 3-layer quality gate, and escalates honestly — built to resolve the repetitive ~78% automatically.
Instant Search
Hybrid semantic + keyword retrieval with cross-encoder reranking over 109 HR docs — cited answers in seconds, in any language. 0.82 Answer Relevancy on Ragas.
Billing Rescue
Rescued a live subscription-billing Telegram bot for a content creator: root-caused a duplicated scheduler granting 60-day subscriptions instead of 30, fixed it under TDD, and killed an expiry job throwing ~880 errors a day.
Listing Pipeline
Automated a property agency's listing workflow: catches private-seller ads, filters out agents, and posts formatted listings to a shared sheet and Telegram channel. Soak-tested: 48 hours, 96 cycles, zero errors.
Anomaly Detector
Three detection layers for month-end close: hard accounting rules, Benford and round-number heuristics, per-client learned models. QuickBooks Online integration, sandbox-tested. 434 tests passing.
What I work with.
Agentic AI Development
Multi-agent systems that plan, delegate to subagents, call tools, and verify their own work.
Advanced RAG Systems
Hybrid retrieval and reranking that answers with the source attached.
Workflow Automation
Repetitive back-office work wired end to end.
Backend & API
The FastAPI services and data layer underneath it all.
LLM Eval & Tracing
Measuring answer quality and watching every call in production.
Prompt Engineering
Prompts tuned and tested against real cases, not guessed.
Orchestration
Retrieval
Evaluation
Serving
Foundation
Who's behind it.
I'm Roman — an AI engineer who ships real systems, not toys. Echofy is a full production SaaS I built alone, from webhooks to formal verification to deployment. Now I help businesses put the same kind of AI to work — fast, and done properly.
Let's put AI to work.
Bring me a problem worth automating. I'll scope it, build it end to end, and ship something that holds up in production.