Selected at 8% acceptance (1.2M KZT scholarship) at ISSAI, Nazarbayev University. Fine-tuned OmniVoice into a description-conditioned Kazakh TTS model: speech is synthesized from text plus a coarse speaker caption (gender + age), with prompt-audio and voice cloning deliberately disabled — it designs a voice rather than cloning one. Trained on the kazakh-speeches corpus (~57.1k train / 3,172 val / 3,173 test, 90/5/5, seed 42) for 5,000 steps on a single H100 80GB: masked-token cross-entropy over 8 audio codebooks, 1e-5 peak LR with cosine decay, bf16, flex_attention, 8,192 batch tokens, grad clip 1.0. Train loss 9.065 → 6.370, best validation loss 6.376 at step 4,500, train/val curves staying aligned with the val curve flattening after ~step 4,000. ~28 min wall-clock at 2.95 it/s. Losses only so far — no human perceptual, intelligibility or speaker-similarity evaluation yet.

1st Place / 190+ teams ($2,500) — Shadow Map AI evolved into Kolenke. AI decision-support for greening, solar & worker safety: Landsat LST heat layer + WebGL shadow simulation (mapbox-gl-shadow-simulator), ranked planting/solar candidates with auditable scoring Priority = 0.55*LST + 0.35*(1-NDVI) + 0.10*exposure, peer-reviewed cooling coefficients (-4°C district cooling, 21kg CO₂/tree/yr). AlemLLM explanations, RAGFlow knowledge base, Kazakh STT voice, Supabase + PostgreSQL/Redis on alem.plus (7 tools). FastAPI + React 19 + MapLibre GL 5.

3rd Place / 70 teams at HackNU. AI banking co-pilot: multi-agent RAG (Supervisor + Local FAISS + Web Search), Gemini 2.5 Flash Live for voice banking, DeepFace/RetinaFace biometric Face ID + JWT auth, predictive spend/budget ML, transaction NLP parsing & marketplace checkout. FastAPI + React Native + FAISS vector DB.

3rd Place / 200 ($2,000). ResNet50 multiclass classifier on 30,000+ images — 92% accuracy, F1=0.895. OpenCV preprocessing + reproducible training/eval pipeline for insurance triage.

Production memory layer for AI agents: ingests turns, extracts structured memories (rule → LLM fallback via OpenAI), evolves facts, pgvector hybrid search + embeddings, canonical slots + superseded history in recall context. FastAPI + PostgreSQL/pgvector + Alembic + Docker, smoke-tested.

Shipped during nFactorial: Instagram parsers (posts/bios/stories/captions) with rate-limit-aware scraping → RAG multi-agent book generation. 800 users, 2,000+ visitors. FastAPI, Redis, async queues, LangChain, Polar.sh payments.