Full-breadth AI Engineer · Agentic systems, deterministic safety layers, production reliability
Based in Finland · Open to freelance, remote roles, and partnerships
I build production-grade systems along two complementary axes — agentic LLM orchestration and disciplined experimental ML research — unified by one principle: a deterministic safety layer wrapped around a probabilistic model.
- LLM orchestration with Claude & the Anthropic SDK — multi-pass extraction, structured-output enforcement, schema validation
- Hybrid systems — probabilistic LLM reasoning paired with deterministic validation to eliminate hallucination, not just monitor it
- Experimental ML research — model calibration, uncertainty quantification, range–null space decomposition; pre-registered, with honest null results
- Event-driven backend (BullMQ, Redis, PostgreSQL) and serverless infrastructure
- API orchestration across heterogeneous sources (REST, GraphQL, WMTS, WFS)
- Accessibility (WCAG 2.2 AA), security hardening, and observability as engineering constraints embedded into system design
No overlays. No superficial fixes. Code-level implementation.
A safety layer for an LLM agent on a critical data path (VAT classification for purchase invoices): the agent extracts facts, a deterministic rule decides, and an uncertain case escalates to a human instead of being guessed at. A 1,000-run determinism check confirms the same input always produces the same decision; four security regression tests cover prompt-injection bypass attempts. Zero dependencies — pure Python standard library, auditable in a single file.
Python (standard library only)
A sibling project to refuse-dont-guess, rebuilt as native Claude Code orchestration — all four runtime primitives in one installable plugin: skill, subagent, MCP server, PreToolUse hook. Live verification surfaced two real bugs, both fixed and re-verified. The bug proved the architecture's core claim in practice: when the subagent hallucinated a figure in its report, the actual VAT decision was still recalculated from ground truth inside the hook.
Claude Code · Python · MCP server · Skill · Subagent · PreToolUse hook
A deterministic layer that labels every pixel of a restored image as measured or invented — separating what the input forces from what the model's prior fabricates (range–null space decomposition). Built as open research: pre-registered hypotheses, locked thresholds, and honest null results. Confirmed finding: calibration slope is content-dependent, verified pre-registered and robust to leave-one-out. One-command falsification test; WCAG AA accessible demo; full research log — including every retraction — public.
Python · PyTorch · ResShift (diffusion) · Range–null decomposition · Pre-registered statistical analysis
An AI assistant built around verifiability rather than raw model output. A LangGraph state machine runs hard-coded guardrail rules after every node, with a hash-chained audit log that detects tampering. Contrasts three perspectives on one synthetic case — traditional manual process, naive AI, and AI with deterministic guardrails — side by side in a Streamlit UI. 20 passing pytest tests, hash-locked dependencies with SHA-256 verification.
Python · LangGraph · Streamlit
A customer-facing digital car salesperson that negotiates price 24/7 — but the negotiation logic is deterministic, not LLM-generated: the dealer configures the acceptance floor and counter-offer steps, the system enforces them, and the LLM never sees or decides the price. Database-level reservation locking (partial unique index, not application-level check-then-act) prevents two customers from reaching payment for the same vehicle simultaneously. A fail-safe external availability check halts the purchase before payment if the dealer's own system can't confirm the vehicle is still available. Includes a multi-tenant DMS import pipeline (dry-run preview → explicit approval → atomic apply) with hash-chained audit trail, and a security review — authentication, output-escaping, and network-isolation fixes verified against live production traffic.
Node.js · Express · PostgreSQL
A deterministic PASS/ESCALATE/BLOCK safety layer (adapted from refuse-dont-guess) running inside a scalable Kafka consumer group, load-tested against a simulated live-TV traffic spike (8,000 msg/s). Peak consumer lag drops from 1,489ms to 400ms as consumers scale 1→4 — measured against the real running stack, not simulated. A WCAG 2.1/2.2 AA accessible dashboard shows the same lag data as a 3D visualization, a reduced-motion 2D gauge, and a semantic HTML table — axe-core clean in both modes. Docker Compose orchestrates 6 services; consumer scaling and Kafka rebalance strategies (KIP-429) run live.
Python · TypeScript · Apache Kafka · Docker Compose · axe-core
A self-directed proof of concept testing an AI-native development workflow against real, unfamiliar legacy code: one representative checkout process from osCommerce, an open-source PHP platform untouched since 2016. A context document mapping business rules was written before any code changes; its first draft got a business rule backwards, trusting a stale 2006 code comment over the code's actual behavior, caught by a failing test rather than a second read. git blame traced the bug to a specific 2015 commit that flipped a condition's logic during a database-layer migration, still live in osCommerce's public master branch ten years later. Four extracted, independently tested units; 34 passing tests; behavior preserved exactly, with the pre-existing bug documented rather than silently fixed.
PHP · PHPUnit · Composer (PSR-4) · Claude Code · Git
Production mobile app and serverless backend orchestrating seven public APIs (SYKE, Finnish Transport Infrastructure Agency, Traficom, FMI, Digitraffic AIS, EMODnet, Cloudflare) into a unified real-time situational picture for boaters. Crowdsourced hazard map with community confirmations, offline-first design (SQLite), GDPR-compliant storage, Cloudflare Turnstile abuse protection.
React Native · Expo · TypeScript · MapLibre · Next.js 15 · Neon PostgreSQL · Vercel Edge
→ frontend · backend · live demo
Production TypeScript WCAG 2.2 AA scanner (~7,500 lines of production code, running on Hetzner) that discovers Finnish business sites, runs accessibility audits, and enriches leads with business-registry data. Claude-powered summaries enable personalised outreach. Hardened infrastructure following a production compromise: Redis authentication, API middleware, SSRF/DNS-rebinding protection, root-execution removal, Vitest coverage, GitHub Actions CI.
Node.js · TypeScript · Playwright · axe-core · Redis · BullMQ · Claude API
Also offering WCAG 2.2 AA accessibility consulting — wpsaavutettavuus.fi
LLM & AI: Claude API · Anthropic SDK · LangGraph · Agentic orchestration · RAG architectures · Multi-pass extraction · Deterministic guardrails · Prompt-injection defense ML research: Model calibration · Uncertainty quantification · Range–null decomposition · Pre-registered design · Statistical analysis Languages: TypeScript · Node.js · Python · PyTorch · React · React Native Backend: Next.js 15 · BullMQ · Redis · PostgreSQL · Prisma · Neon serverless Infrastructure: Vercel Edge · Docker · Hetzner · Linux · Cloudflare · OAuth 2.0 · System hardening Accessibility: WCAG 2.2 AA · ARIA · Semantic HTML · NVDA/VoiceOver testing · Mobile accessibility (React Native)
LLM systems fail in production when reliability, validation, and operational constraints are treated as afterthoughts. I work where probabilistic reasoning meets deterministic logic — where hallucination must be eliminated, not just monitored; where accessibility is engineered into the architecture, not audited at the end; and where the same discipline applies to my own research as to my code: pre-register before the data, flag what's uncertain, retract what doesn't hold.