06 / INTELLIGENCE
Tornike Kalandadze
AI Product Engineer & Technical Lead
LLM systems · Voice agents · Applied ML · Guardrails
Profile
AI product engineer and technical lead building complete user-facing systems around models—not model demos. I lead AI Academy’s four-person product team and delivered Mia’s Georgian voice-agent stack. My approach combines structured outputs, grounded context, explicit tool boundaries, cost and usage visibility, deterministic fallbacks, review gates, and human control over consequential actions.
Experience
Lead Developer & Partner
AI Academy
Apr 2026–Present
Launched AI product · four-person team
- Lead the product and engineering of a Georgian-native learning platform that turns a syllabus into structured lessons, quizzes, and a grounded Socratic tutor.
- Designed a three-stage Gemini generation pipeline, Claude streaming tutor, Zod-validated structured outputs, review gates, usage/cost logs, and abortable cleanup.
- Own architecture, task direction, code review, deployment, and product decisions through public launch and advertising rollout.
AI & Mobile Engineer
Mia · client work
2026
Signed Android release
- Built the Georgian voice loop across on-device wake-word inference, Google Chirp 2 speech recognition, OpenAI tool calling, sentence-level SSE streaming, and ElevenLabs speech.
- Integrated real mobile actions—alarms, timers, weather, music, calculations, and Georgian/Russian/English interpretation—with background-safe playback fallbacks.
Independent Applied-AI Engineer
Selected prototypes & hackathon work
2026
Local ML, explainable scoring, human gates
- Applied IsolationForest anomaly scoring in Sentinel with a deterministic baseline and rule-based fallback; designed every high score to wait for human judgment.
- Built ShieldMesh solo and won the VibeCoding From 0 hackathon with an offline-first threat-intelligence system combining on-device detection, mesh relay, and Solana verification.
Selected engineering work
AI Academy
Lead developer & partner
Georgian-native platform that converts source material into coherent lessons, quizzes, spaced review, and grounded tutoring.
- Gemini parses and expands syllabi; Claude streams lesson-grounded tutoring and grades quizzes behind a provider abstraction.
- Zod validates every model boundary; malformed output retries, generated lessons pass a review gate, and token usage/cost is logged per call.
Claude · Gemini · Next.js · Zod · Supabase
Mia
Client work · voice agent
Android-first Georgian assistant connecting local wake-word inference to cloud speech, tool use, and natural voice output.
- Google Chirp 2 transcription → OpenAI tool calling and streamed response → ElevenLabs speech, with sentence-level delivery over SSE.
- Wake word stays on-device through ONNX; tools perform alarms, timers, weather, music, calculations, and two-way interpretation.
OpenAI · Google Chirp 2 · ElevenLabs · ONNX · React Native
Sentinel
Applied-ML prototype
Local privacy watchdog using anomaly scoring to rank OS activity without outsourcing personal event data to the cloud.
- IsolationForest scores event rate, actor novelty, target sensitivity, and time of day; a deterministic baseline avoids an empty cold start.
- The ML dependency is optional: C# rules preserve core behavior, and alerts wait for human review.
scikit-learn · Python · C# · SQLite · WPF