I amIvetta Dashkova

A single Next.js + Postgres app where four product modules share one geospatial datastore (PostGIS for geometry, pgvector for semantic search, tsvector for keyword). It showcases end-to-end AI orchestration — a streaming tool-calling agent (also exposed as an MCP server), hybrid RAG with a grounding gate, and structured LLM output — alongside real geospatial engineering (PostGIS trajectory analytics, marine routing, animated map playback). This is the reusable geo-AI template that Domus (below) productizes for real estate.

• Architected a single-database platform — one Postgres with PostGIS + pgvector + tsvector powering 4 modules, with no syncing between a separate vector DB and a geo DB.
• Built agentic AI orchestration: a streaming, multi-step tool-calling agent over 7 real geo tools, exposed both in-app and as a Model Context Protocol (MCP) server — one tool core, two surfaces.
• Engineered geospatial analytics in PostGIS: stay-point stop detection, elevation-gain hysteresis, sea-route pathfinding for ships, and time-based animated map playback.
• Implemented hybrid RAG (vector + keyword + spatial) with reciprocal-rank fusion and a grounding gate that declines out-of-corpus questions — cited answers, no hallucinations.
• Instrumented the whole pipeline with Langfuse + OpenTelemetry and a cross-module eval harness — LLM quality tracked as a first-class metric.

A full-stack English-learning platform — a React + Redux SPA backed by a Node.js / Express + MongoDB REST API — where AI does the heavy lifting. A streaming, tool-calling agent reaches into the learner's own data (weakest words, dictionary, answer grading) so guidance stays grounded and per-user; an enrichment agent turns a bare word into a full card via schema-validated LLM output; and a Stripe credit paywall gates the AI. Layered on top: gamified practice, a custom SVG analytics dashboard, JWT auth, a test suite and full technical SEO. Designed, built, documented (OpenAPI/Swagger) and shipped to production solo, on a near-free stack.

• Full-stack, solo on a different stack from my geo-AI work: a React + Redux SPA and a Node.js / Express + MongoDB REST API, with JWT auth, server-side sorting / pagination and rate limiting.
• Built a streaming, tool-calling AI tutor that queries the learner's own words and grades answers server-side — grounded, per-user, no hallucinations.
• Added a dictionary-enrichment agent: schema-validated structured LLM output (translation, part of speech, IPA, examples) that drops straight into the add-word form.
• Integrated real Stripe payments — a credit paywall with an embedded Payment Element, webhook fulfillment and an idempotent verify-on-return fallback.
• Engineered a gamified learning engine: 6 exercises per item (type / choose / gap-fill × EN↔UA), monthly spaced re-review, and an analytics dashboard of hand-built SVG charts.

The real-estate productization of the Locus geo-AI template (above): a multi-tenant operations tool for property agencies. It turns a free-text buyer enquiry into grounded property matches, plans an optimized viewing route across a city, and values a listing from comparable sales — every AI answer backed by a retrieval or tool call against the agency's own catalog, prompt-injection-resistant and never free-form. What it adds beyond the template: hard multi-tenant isolation, road-network routing, an automated valuation model and a production ingest pipeline — built and shipped to production solo, on open data.

• Enforced multi-tenant isolation with fail-closed Postgres Row-Level Security (audited in CI) and an idempotent pg-boss ingest pipeline (geocode → embed → enrich) with exponential-backoff retries.
• Engineered a viewing-route optimizer: a TSP over self-hosted OSRM driving times, with a straight-line fallback transparently flagged when a stop falls outside the routing region.
• Implemented an automated valuation model (inverse-distance-weighted comparables) with confidence scoring, validated by leave-one-out MAPE.
• Ran the whole retrieval stack on local, zero-cost embeddings — Transformers.js bge-small for text and CLIP for image search — so nothing leaves the free tier.
• Fused semantic (pgvector), keyword (pg_trgm) and spatial (PostGIS) ranking with reciprocal-rank fusion, measured offline with precision@k / recall@k / MRR.