Hands-on CTO at Livable Places

At Livable Places I did not work on a generic "AI-first" story. I worked on a real delivery system for a proptech product with actual operational responsibility. That spanned monorepo product development, distributed scoring and data workflows, self-hosted platform operations, geo/census services, and the evidence-led preparation of a new office-risk product line.

The product thesis is bigger than a score dashboard: make societal demand for real estate uses measurable and comparable at location level. The operating logic is Demand → Supply → Property, so financing, acquisition, ESG, and portfolio decisions can rely on more legible signals.

My leverage was connecting product logic, data pipelines, AI/ML research, geo services, queue orchestration, and platform operations instead of treating them as separate workstreams. The monorepo, workers, auth, observability, GitOps-style deploys, backups, and runbooks form one system that can be built, shipped, and operated. Collaboration was distributed, including design capability from Ukraine.

The AI/ML work stayed deliberately evidence-bound. Office-risk fingerprinting, PCA, k-means, Location Absorption Score, benchmarking, and feature ablation are used for characterization and screening until target quality, residualization, and cross-city stability can support stronger claims.

From late 2025, most of my implementation was produced through agent workflows. My work shifted from writing most code line by line to orchestration: framing outcomes, context, and constraints, directing agents, reviewing changes and evidence, and remaining accountable for product and technical decisions.

• Built and evolved a pnpm/Turborepo monorepo for the web app, services, workers, and operating tools
• Testing base: used Cypress briefly and then moved to Playwright as the primary browser/E2E testing stack
• Worked hands-on on distributed scoring and data workflows with Bun, BullMQ, FlowProducer, QueueEvents, and Redis
• Used Redis for locking, Pub/Sub, and job-completion signaling instead of fragile glue logic
• Built authenticated async report export paths with service-side auth guards, shared export domain logic, and XLSX generation
• Owned a self-hosted bare-metal platform with Docker Swarm, Traefik, Authentik, Ansible bootstrap, Docker Secrets, restic backups, WireGuard, and a full observability stack
• Established IaC/GitOps-style GitHub workflow paths for build, release, deployment, rollback, scaling, backup, restore smoke tests, maintenance, and self-healing
• Worked with distributed external roles, including a designer in Ukraine, through explicit artifacts, review paths, and visible delivery goals
• Worked hands-on on the geo/census service layer and shaped its architecture around Kotlin, Spring Boot, PostGIS, OpenAPI, FastAPI, Valhalla, OSM, and external POI imports
• Personally built a Python data pipeline for external doctor and school data: S3/CSV ingestion, validation and normalization, retry logic, temporary tables and bulk import into PostgreSQL/PostGIS with JSONB tags, tests, containers, and GitHub Actions
• Prepared a new office-risk product line through evidence-led product and delivery work
• Designed and challenged ML workflows around PCA, k-means clustering, LAS calibration, benchmark matrices, target quality, residualization, and feature ablation
• Used multi-agent research workflows and multi-agent pipelines (asset-research and factsheet-review skills, observation → inference → hypothesis → experiment) as part of the product evidence layer, not only as developer convenience
• Since late 2025, delivered most implementation through Claude Code, Codex, Cursor, Antigravity, OpenCode, and Pi Coding Agent; outcomes, context, constraints, review, and evidence gates remain my responsibility, complemented by local agent skills and plugin-style skill development
• Translated strategy into visible goals, scopes, work items, development boards, releases, demos, and customer-facing feedback loops
• Covered S-Core product clarification and story writing for several months without an assigned Product Owner
• Built demo and campaign measurement paths with anonymous access, UTM tracking, event analytics, and conversion-funnel thinking
• Built internal planning and reporting tooling with Nuxt, Vue, SQLite, Zod, and localized product surfaces

• Delivery as a system: monorepo, platform, queues, geo services, Cypress-to-Playwright test migration, and product work are not optimized in isolation
• Decision quality over dashboards: product work is judged by whether it helps customers make better financing, acquisition, ESG, or portfolio decisions
• Infrastructure as code: platform state, auth baselines, stack definitions, secrets, backups, and operations are encoded and repeatable
• Operability by design: logs, metrics, traces, alerts, and auth are part of the product, not follow-up work
• Distributed orchestration with explicit dependencies, locking, and visibility instead of silent background jobs
• AI-first where it matters: use agents, LLM tooling, and research automation to improve evidence quality and delivery speed, but keep claims bounded by data
• ML with evidence gates: fingerprint, benchmark, calibrate, residualize, ablate, and only then claim predictive value
• Geo/data products as infrastructure: OSM, census, POI, routing, and market data are treated as versioned, observable product dependencies
• Visible operating model: strategy, roadmap, goals, scope, work items, board, release, demo, and customer feedback stay connected
• Small batches and visible ownership instead of roadmap and process signaling
• Evidence-first product work: observation, hypothesis, experiment, then scale