M. Tarik Altuncu, PhD

London

Joined at a small-team stage in May 2025 and have shipped through several iterations of the platform. The current 6-person engineering team (3 ML/AI engineers, 1 iOS, 1 designer, 1 founder/PM) came together over the second half of 2025.

Personalised immersive language acquisition platform: multi-persona voice agent tutors, AI-generated articles, FSRS-based vocabulary scheduling.

• Multi-provider STT and audio-LLM evaluation harness. 30 model/method combinations across OpenAI, Gemini, Mistral, and ElevenLabs. 1,500 API calls on real production audio. Consensus-based WER with 95% CIs and p95 latencies, plus Pareto-frontier analysis. Found Mistral Voxtral Mini STT outperforms GPT-4o-transcribe on learner speech at lower cost; the harness doubles as a regression test against vendor model updates.
• TTS provider selection. ElevenLabs Multilingual v2 for conversational tutor personas; Cartesia Sonic 3 for article read-out, where reliable streaming word-level timestamps in non-English languages outweighed raw quality.
• Three-layer voice context injection. Gives the stateless real-time tutor working memory of every prior session. Mem0 with event-anchored timestamps and expiration-aware future facts. Each layer degrades gracefully when an upstream service is unreachable.
• Personalised push notifications. Body copy generated per user by Claude Sonnet from the recommendation feed, timezone-aware APNs delivery, and a content-freshness check that drops the push if the user already saw the content in-app. 5.3x lift in notification-driven opens within the first week of full rollout. 7-state lifecycle across 4 services and 2 Redis queues.
• On-demand article streaming with word-level audio alignment. Built jointly with another backend engineer. SSE from partial-JSON LLM output; sentence-buffered TTS with shifted timestamps yields a contiguous read-along timeline. Generate-once-per-shared-article semantics via a Redis claim.

Supported alongside the team (primarily owned by the other backend/AI engineers): the Pipecat + LiveKit + Modal voice cascade (sub-second turn latency) and the hybrid recommendation engine with two-layer diversity enforcement that addresses LLM topic convergence (the "coffee problem").

Stack: Python, FastAPI, Pipecat, LiveKit, Modal, PostgreSQL, Redis, Mem0, Latitude, OpenAI Realtime, Claude, ElevenLabs, Cartesia, APNs.

San Francisco

Highlights of two years on the AI team:

• Retrieval-Augmented Generation system for in-product knowledge search. Among the early enterprise RAG deployments at the time. Milvus and ElasticSearch hybrid retrieval, qualification gate for generated answers, tenant-isolated prompts. Recognised as one of the differentiators in the product's path to the Gartner leadership quadrant.
• AI-powered virtual assistant chatbot from POC to production. Presented strategy options directly to the CTO.
• Multi-tenant collaborative-filtering recommender on Snowflake, Airflow, MLFlow, Redis, and Kubernetes. Over 100K daily recommendations served.
• Fine-tuned Microsoft Phi for personalised content relevance classification at batch-inference scale. 79% drop in false positives.
• Customer churn model with LightGBM on Snowflake data. 87% recall at 60-day lead time; ~$5M ARR in potential save.

Designed and taught two new courses for the Computer Engineering programme: an Artificial Intelligence course and a Generative AI course (cmp405genAI/cmp405). Currently on unpaid leave to focus on Amble.

Founders Inc. Winter Cold Start accelerator (San Francisco). Designed a multi-modal AI pipeline combining LLMs, image generation, and audio synthesis for a personalised children's storytelling app. Soft-launched in Middle Eastern markets.

• Global Discovery RAG news system (TRT). LangChain and LangGraph pipeline for a global newsroom with editorial guardrailing and an intent-classification evaluation framework.
• MediSummarize clinical documentation assistant. Unsloth-based LLM fine-tuning on medical datasets for privacy-preserving clinical-note summarisation. 63% reduction in physician documentation time.

Set the data and analytics direction for a public broadcaster's OTT media organisation, leading a 17-person team across Growth Data Science, Data Engineering, and Machine Learning. The platform reached 5M+ video, 600K+ audio, and 2M+ mobile-game users during this period.

ML Team Lead at TRT (2021, 7 engineers across NLP, vision, personalisation; MLOps on DVC, MLFlow, Airflow). Senior Data Scientist at TRT (2020 - 2021, real-time pipeline on 6M+ daily tweets; transformer sentiment for TV drama). Data Scientist at TRT World (2016). Social CRM lead at Adba International (2014 - 2016; Salesforce/Brandwatch/ClaraBridge partnerships for Turkish Airlines and others).