Sezgin Saygili

Patientdesk

I build AI phone agents and the web app clinics use to manage them. The agents answer calls, book and reschedule appointments, check insurance and take payments through integrations with clinic software, including OpenDental, CareStack, CorePractice, Dental4Windows and NexHealth. My work includes call logs and analytics, keeping each clinic's data separate, and letting clinics set up their agents themselves.

TypeScript · Next.js · Python · FastAPI · Supabase · PostgreSQL · GCP · LiveKit · ElevenLabs · Stripe

Alania-2 & Duyu

I trained Alania-2, a 488M-parameter Turkish text-to-speech model, from scratch on about 18,000 hours of speech and launched it at speech.patientdesk.ai. It generates 48 kHz audio through OpenAI-compatible REST and WebSocket APIs. Our published measurement reduced median time to first streamed audio from ~365 ms to ~190 ms, including the network, on 5 October 2026.

I built the serving platform, browser playgrounds, API-key and usage workflows, and Turkish text normalization. Alania-2 and the legacy model share GPUs through NVIDIA MPS. Duyu adds file and live transcription under the same key. Style instructions and voice design are previews; consent-gated voice cloning is in beta.

Python · FastAPI · PyTorch · CUDA · NVIDIA MPS · WebSockets · TypeScript · Next.js

Antalia-2 Mini

I released Antalia-2 Mini, a 7.62M-parameter Turkish TTS model with 48 kHz output, under Apache 2.0. It includes a Python package and CLI, sentence-chunk streaming, and ONNX inference entirely in the browser. It is also available free through the speech API. One synthetic voice; no cloning.

Our published CPU benchmark measured about 50× real-time synthesis on an Apple M5. The model recorded about 1.1% word error rate on the 495-sentence Freya-TR-Eval comparison, using Whisper transcription and three takes—not a human listening score.

PyTorch · Flow matching · Vocos-style vocoder · ONNX Runtime · Python

ImpliedOptions

I built ImpliedOptions for web, iOS and Android: options flow, volatility analysis, strategy modelling and alerts. The web product now includes Theo, an AI research assistant, and 23 tools available through REST and MCP, the protocol used by AI assistants. Users can research trades, manage watchlists and model positions with up to eight option legs. The Python data services ingest trades, identify blocks, sweeps and multi-leg strategies, and store historical data in PostgreSQL.

Python · FastAPI · PostgreSQL · Docker · Next.js · React · Flutter · MCP

VoiceData

I built VoiceData, a platform for paid voice recording and audio review. Contributors apply, submit recordings and receive payouts for approved work. I built the recording, quality checks, review and payout workflows, and used the platform to collect human judgments for the AI Voice Benchmark.

Next.js · Supabase · LiveKit · Python

Open Turkish speech data

I released 1.44 million generated Turkish sentences, 3,411 hours of synthetic speech across 2.04 million clips, and bilingual style annotations for 835,267 real-speech segments. The releases include provenance, quality metadata and pronunciation corrections. The style dataset references source audio rather than redistributing it.

These three Alania datasets had 6,079 combined lifetime Hugging Face downloads as of 11 October 2026. Across all six public datasets and six models on my profile, the total was 9,185. Download counts are not unique users.

Python · PyTorch · vLLM · Hugging Face Datasets · Parquet

News relevancy classifiers

I fine-tuned two BERT models to score news headlines on a four-point relevance scale. FinBERT-ft-v3 uses about 225 hand-labelled Finnish headlines; BERT-ft-v2 uses about 254 English healthcare and AI headlines. Both are published on Hugging Face as small research models, not general-purpose classifiers.

Python · PyTorch · Transformers · Hugging Face