Tazo Lezhava

Quissly is an AI product discovery platform that helps online retailers show shoppers the right products. As co-founder and CEO, I own the technical architecture end to end and built the core AI stack: hybrid keyword and semantic search over merchant catalogs, multimodal embeddings across product images, text, and metadata, LLM re-ranking that scores results against the shopper's full request, and a conversational shopping agent that keeps context across turns. Merchants install it through a self-serve Shopify app. The platform serves 50+ retailers across 12 countries. I also built the company's internal AI systems for growth: outbound agents, a data enrichment pipeline, and prospect scoring. The longer-term aim is catalog infrastructure ready for an agentic world, where AI agents increasingly browse and make purchase decisions on shoppers' behalf.

homeview.ai is an AI-powered interior design SaaS with built-in product recommendations. As co-founder and CTO from 2022 to 2024, I was responsible for building the product: I set the technical architecture, led development of the AI design features, and built the recommendation layer that connects a user's design to real products they can buy. It was the first of two AI startups I have co-founded, followed by Quissly.

At TBC Bank, the largest commercial bank in the Caucasus region, I built the data science function and led the development of predictive models that informed the bank's core decisions. These included credit underwriting and scoring, IFRS 9, fraud detection, anti-money laundering, churn, customer lifetime value, and a Next Best Offer engine that ranks each customer, product, and channel combination by expected value. I moved the bank from simple rule-based and linear models to modern machine learning, introducing Databricks, Azure, AWS, MLflow, and AutoML to enable weekly retraining and model improvement. The models ran as a separate ML service that fed the bank's legacy core systems via APIs, and the credit and Next Best Offer models delivered significant, measurable revenue gains.

As head of data science and AI at TBC Bank, I led the bank's generative AI work and the governance around it. My team fine-tuned LLMs and built local Georgian-language models for customer-facing chatbots and an internal knowledge-base assistant, keeping customer data within the bank rather than sending it to external APIs. I also founded the bank's Model Risk Management function, the independent supervisor for every model across TBC Group, and wrote the bank's MLOps and LLMOps standards. I defined which AI tools staff could use and what data could be used in them, made MRM validation mandatory before any LLM launch, and required human review of customer-facing AI output.

Geniti is a platform for solo founders to build and run AI organizations. Instead of hiring a team, a founder works with a persistent AI Co-CEO that understands the venture, breaks goals into tasks, and delegates them to specialist agents. These include a Researcher who conducts web research and writes structured competitive reports, and a Content Specialist who drafts communications in the founder's voice. I built it solo with Claude Code: a Python/FastAPI back end, a React/Next.js front end, and GCP infrastructure (Cloud Run, Cloud SQL). The platform routes work across multiple frontier models (Claude Opus, Sonnet, and Haiku, and Gemini Pro and Flash), uses Tavily for web search, and supports MCP servers for tools.

Gipity is a safe AI playground for children aged 8 to 14, built as a policy-driven gateway between kids and large language models. An LLM classifies every message in real time by topic and risk level, and that drives a parent-defined policy engine: encouraged themes, strict blocklists, and tone settings (patient, encouraging, or Socratic). If a conversation touches a dangerous topic such as bullying or self-harm, the AI is instructed to stay supportive and de-escalate, and the parent gets a high-priority alert. A parent portal displays a child's interests and themes without parents having to read every message, and transcripts are automatically deleted after 7 days. As founder, I designed the product, the safety taxonomy, and the back end: Python/FastAPI on Google Cloud Run, with IAM/OIDC authentication, and integrated with OpenAI, Gemini, and Claude models. The back end is live, and the iOS app is in development.

FIT (Fast Insight Testing) is a research tool I created within BlackRock's Systematic Active Equity group to quickly evaluate alternative data sources. Given a single candidate data source, FIT automatically generated thousands of alpha signals from it, backtested them, and produced interactive dashboards. Portfolio managers and researchers used those dashboards to explore the results and make informed decisions about which data to bring in. FIT supported SAE's Data Search Team, where I was one of three members responsible for sourcing and onboarding new data across the group's global strategies. Built in Python.

Fins is a web app that a two-person household uses (for now) to track and plan money without bank connections. Transactions arrive by manual entry or by importing statements as CSV, Excel, or PDF, and a deterministic rules engine classifies them by description, amount, direction, and merchant category code. Gemini on Vertex AI performs three tasks: extracting rows from PDF statements, proposing rules for uncategorized transactions, and mapping spreadsheet headers to import fields. It never classifies a transaction itself. I also built nested categories, budgets, balance snapshots, currency conversion from National Bank of Georgia rates, and email reminders.

The stack is Python 3.12, FastAPI, Jinja2, HTMX, and Alpine.js, SQLAlchemy 2 and Alembic over PostgreSQL, on Cloud Run with Cloud SQL. GitHub Actions deploys via Workload Identity Federation, gated on a 1,486-test pytest and Playwright suite.

I set the architecture, wrote the specs, and built much of it with one other developer. All milestones are deployed.