The model is just a tool. The edge is understanding people.
I'm an LLM application engineer: I don't train models from scratch — I integrate them with real systems: backend, database, authorization, deployment, cost and monitoring. Behind me are fourteen years of commercial software engineering, from e-learning and banking-sector systems to enterprise front-end in international projects for aviation, e-commerce and pharma.
The second pillar of my profile is unusual: two master's degrees — in developmental psychology and pedagogy — and five years of university lecturing. That's why I specialize in human-in-the-loop systems: ones where AI supports a person's decision instead of making it for them. I know why people distrust systems built to "help" them — so I design AI that supports a human decision instead of quietly taking it over.
I built every system below myself and use them daily. This is deliberate dogfooding — I carry the cost of design mistakes on my own skin, before anyone else does.
Trader AI Assistant
Mentoring platform · multi-agentA production platform where an AI mentor guides the user according to their own written methodology — instead of generic advice. Multiple cooperating agents on the Claude Agent SDK, prompts assembled dynamically from the methodology and user history, integration with a brokerage account (auto-journaling, a live risk-rule guardian). The system runs on real money — AI mistakes carry a real cost here.
MORE — ARCHITECTURE & COMPETENCIES
- End-to-end architecture. React 19 + TypeScript SPA (Vite, MUI, Zustand), Firebase Auth/Hosting/Storage, Express + TypeScript containerized on Google Cloud Run, PostgreSQL on Supabase. All domain data flows through the backend API; model API keys live server-side only.
- LLM layer with two providers (Anthropic Claude and Google Gemini) and backend-level dispatch. Every prompt is assembled from blocks: trader profile, session history, market context and the user's own written methodology — setups, checklists and risk rules are data in the system, so the mentor enforces the user's rules, not generic advice.
- Multi-agent layer on the Claude Agent SDK running as background jobs, with an isolation rule I designed: an analyzing agent cannot see the mentor's earlier assessments before issuing its own.
- Model ops. A central model registry in the database (engines swappable without a deploy), request limits per subscription plan, BYO API keys, and a cost-driven default-model choice — comparable quality at a fraction of the price.
- Production discipline. 100+ numbered SQL migrations, Vitest unit tests, Playwright E2E on ephemeral accounts, and procedures forged on real incidents — e.g. an env-replacing deploy flag led to a merge-and-verify procedure; a broker API returning HTTP success on rejected operations taught me to validate response bodies, not status codes.
Deep Mentor
Personalized retrieval (RAG)The mentor's long-term memory layer: the model reads the user's personal knowledge base — notes, lessons, hypotheses — and cites it in its analyses, instead of answering from general knowledge. Base seeding, digest in prompts, hard control over inference cost.
MORE — ARCHITECTURE & COMPETENCIES
- Per-user wiki memory in the database: agent pipelines analyze the user's day and write notes into their personal knowledge base, which the mentor then cites in its analyses.
- Retrieval designed for cost. Base seeding, digests in prompts instead of raw dumps, expensive deep runs only on demand — inference cost is a first-class design constraint, not an afterthought.
- Two feed modes. The knowledge base can be filled by the platform's own automations or by an external pipeline compiling and pushing notes from the user's local Obsidian vault.
audit-mentor
LLM quality evaluationA harness for the cyclical audit of the AI mentor's answers: an error taxonomy in four classes, deduplication of findings by root cause, a remediation backlog with locations in the code. Organizing principle: deterministic code computes the metrics, the model only interprets — an LLM doesn't grade itself on data it could distort.
MORE — ARCHITECTURE & COMPETENCIES
- Cyclical audit (~every two weeks) of the mentor's production answers. Every finding carries an evidence quote and a root-cause hypothesis — no vague "could be better" notes.
- Regression backlog. Findings are deduplicated by root-cause signature, an occurrence counter works as the priority weight, and each entry points to path:line locations in the codebase.
- Division of labor by design. Deterministic code computes the metrics, the model interprets them, and the human reviews findings after the fact and kills false positives.
- Pulled out into an open-source library. The same lifecycle showed up on its own in a second system of mine, so I moved it into a Python library: dedup by root cause, occurrence counters, regression cases written before the fix, false positives retracted instead of deleted. It is the part that prompt-test runners and observability platforms leave to you. The ledger stays a markdown file in git that a person owns — the library only edits it in place, and its adapters read another tool's output file rather than importing its SDK. Pre-release, ready to publish.
Prompter & Murmur
Realtime voice AI · macOSNative apps for live conversations: streaming transcription of both sides, simultaneous translation, contextual prompts pulled from a personal knowledge base — and speech synthesis that answers in the user's own voice, running locally. The main engineering challenge: the latency budget of the audio → text → LLM → response pipeline, across two parallel model sessions.
MORE — ARCHITECTURE & COMPETENCIES
- Streaming transcription of both sides of a conversation with two model sessions running in parallel; the core engineering work is the latency budget of the audio → text → LLM → response pipeline.
- Hybrid speech stack. On-device Apple SpeechAnalyzer where the language is supported, Gemini Live where it isn't — chosen per language after real spikes on the target OS, not from documentation alone.
- Contextual prompt cards retrieved live from a personal knowledge base during the conversation; simultaneous translation for cross-language calls.
- It speaks, too — in the user's own voice, entirely on the machine. A locally run voice model reads the translated turn back in a clone of the user's voice, so a cross-language call still sounds like one person talking instead of a robot standing in for them. Nothing leaves the laptop and there is no per-character bill. The hard part was measurement, not prompts: the reference recording had to drop below twenty seconds because the library was quietly trimming longer ones, and the standard similarity scores kept praising takes that sounded worse to the ear — so the ear became the acceptance test and the metrics were demoted to a filter for outright failures.
Scout
Research agent · MCPA market-analysis agent with a hard autonomy boundary: it analyzes, suggests and alerts, but the decision always stays with the human — zero auto-trading, no exceptions. Deterministic metrics plus a narrative layer from web research, a walk-forward backtesting lab, tool integrations over the Model Context Protocol, including a bridge to TradingView Desktop.
MORE — ARCHITECTURE & COMPETENCIES
- A market radar that separates roles: deterministic code computes volatility metrics, a narrative layer from web research adds context — the model never invents the numbers it reasons about.
- "Blind" analyses. The agent produces its analysis without seeing the human's view, and only then are the two diffed — a learning loop deliberately designed to avoid anchoring.
- Setup Lab: walk-forward backtesting of strategies with an append-only experiment journal; tool integrations over MCP (including a bridge controlling TradingView Desktop) plus direct exchange REST APIs.
Python trading system
~35K lines · case studyA production asynchronous service: a layered architecture with the execution service as the sole owner of orders, centralized risk configuration, unit tests on the money path, demo mode as the safe default. Alongside it — a signal engine with LLM-based signal scoring in a live loop. Private code, discussed as a case study in interviews.
MORE — ARCHITECTURE & COMPETENCIES
- ~30.7K lines: an async FastAPI service with a layered architecture — connectors → analysis → an execution service as the sole owner of orders. Centralized risk configuration, demo mode as the safe default.
- Tests where money moves. Unit tests concentrated on the money path: loss limits, pause-after-loss, signal-age validation.
- ~5.1K lines alongside: a 12-rule volume-price signal engine with a live candle-by-candle loop and LLM-based signal scoring; SQLAlchemy/Postgres persistence.
Atelier
AI in professional development · EdTechA year-long professional-development program for someone entering the styling profession — led by an AI agent under my subject-matter supervision. A daily newsletter generated from web research and the participant's personal profile, audio lesson series released one per day, a dashboard with the day's task, and a feedback loop: the participant's entries feed back into the generator of the next issues. The psychology of adult learning translated directly into system architecture.
MORE — ARCHITECTURE & COMPETENCIES
- A headless agent on a daily cron generates each issue from web research, the participant's profile and the program's weekly themes; an anti-repetition rule makes it read the last 14 issues before writing a new one.
- Audio lessons and a dashboard. TTS lesson series released one per day from a queue; a dashboard with the day's task; the participant's entries and check-ins feed back into the next issues.
- Curriculum built on psychology. A year-long program map designed around the psychology of entering a new profession — dosing, celebrating streaks, support instead of accounting.
A lot of what I have been doing for years goes by the names of libraries I have never touched. If one of those names comes up, I say I don't know it. Then I take a project that needs it and learn it by building.
I detected sycophancy in my own AI mentor: it anchored on the user's emotional self-assessment and confirmed it, instead of confronting it with data. As a psychologist I know the difference between supporting and agreeing — which is why a mentor that doesn't just agree has to be designed.
The most important architectural decisions are about what the system must not do. Zero auto-trading in the market agent. The only mutation on the broker's side is a protective daily lockout. Designing such boundaries requires understanding a person under pressure — their fatigue and their impulses.
Deterministic code computes the metrics, the model interprets, the human reviews after the fact. A deliberate human-review-loop architecture — instead of trusting the model to grade itself.
I'm looking for projects at the intersection of AI engineering and people: mentoring and educational products, human-in-the-loop systems, LLM applications from architecture to production.
B2B contract, fully remote. I reply to every concrete message.