Applied AI Engineer | Python, LLM applications, and workflow automation
I build AI tools for people working with documents, knowledge, and recurring decisions. My projects combine Python backends, retrieval, external integrations, and human review.
I start with the existing workflow: what takes time, where mistakes matter, and which decisions should stay with a person. Then I build and evaluate a focused solution, using deterministic code and language models where each fits.
Based in Tbilisi, Georgia. Open to remote Applied AI Engineering and AI Solutions Engineering roles.
| Project | Who it helps and what I built | Current stage |
|---|---|---|
| gdev-agent | Python/FastAPI backend for player-support triage, with signed webhooks, human approval, tenant isolation, and audit trails | Local reference implementation; conformance passes, harder challenge gate fails |
| Telegram Personal AI Assistant | Source-backed research, briefs, monitoring, and confirmed actions in one Telegram conversation, with durable state | Implemented prototype; connected-account pilot and daily operation unconfirmed |
| Dream Motif Interpreter | A practitioner's dream-archive workflow: Google Docs ingestion, search, voice input, notes, and reviewed motif suggestions | Single-operator prototype; public evidence uses synthetic records |
| AI Workflow Playbook | My delivery method for developers and workflow owners: frame the problem, choose a change, implement, verify, and hand over | Public reference material; Product/Desktop tools in private evaluation preview |
| Eval Ground Truth Lab | Python CLI for comparing model, prompt, tool, and policy changes against versioned cases and quality, cost, and latency gates | Released local alpha |
| Agent Runtime Grid | Experimental Python job runtime with PostgreSQL state, Redis queues, retries, cancellation, and reproducible failure tests | Local runtime experiments |
Run a regression-and-fix demo: follow an approval-policy regression through detection and a passing fix, then inspect the reports and remaining human decision.
- Georgia Community Navigator — Telegram topic guides and source-backed assistance for administrative tasks in Georgia. Guide/navigation and opt-in AI work are implemented on the development branch, with bounded S1 semantic acceptance recorded; live testing and release remain unperformed. The delivery order is useful guides and onboarding, paid assistance, then the specialist foundation. Private source repository.
- English Teacher Assistant — discovery and technical design with an English teacher: reuse teaching materials, prepare faithful homework previews, revise by text or voice, and publish after teacher approval. Booking and administration form a parallel workflow. Application implementation and real-account integrations have not started. Private source repository.
- Python services and integrations: FastAPI, PostgreSQL, Redis, Telegram, and document workflows.
- Retrieval and source handling: ingestion, hybrid search, versioned annotations, provenance, and human review.
- Evaluation and debugging: regression cases, deterministic checks, failure analysis, and explicit quality, cost, and latency boundaries.
- Workflow discovery: translate a domain expert's working process into a focused implementation and a reviewable result.
For a deterministic case, see Trader Risk Audit: trade CSVs, explicit rules, and traceable reports in a local Python CLI. Workflow-to-Agent Studio shows the earlier design step, from SOPs and interviews to reviewable implementation blueprints.
Before software engineering, I worked as an international technical educator at Schwarzkopf Professional. That experience shapes how I clarify requirements, explain technical choices, and collaborate with domain experts.
Project stages reviewed on 24 October 2026; each repository retains its detailed evidence and limitations.