Speak · 2022–25
Product & AI engineering
As the LLM lead, I worked closely with Engineering, Product, and Design (EPD) to research, prototype, architect, and ship a number of AI initiatives across the product. My focus was on getting learners to speak as much as possible and then generating personalized curriculum based on their individual patterns of mistakes.
Conversational Lessons 2022
Prototyped ASR → LLM → TTS conversational lessons to help learners practice their English in numerous real-world scenarios, and provided them with feedback on their speech. I tech-led Python/FastAPI backend, API contract, DB architecture, and rollout; also modeled inference economics, pricing, and GPU capacity. This became the foundation for a new Premium Plus subscription tier and meaningfully increased ARR.
At launch: 40K lesson starts/day · 700 RPM · 500K TPM
Lesson Reviews 2023
Built a personalized lesson service that ingested a full conversational lesson, extracted and ranked errors, and instantly generated a follow-up lesson for users to review their most important errors. Tech-led the entire project, and PM’d the iteration into a default end-of-lesson flow increasing lesson finish rate and subscription upgrade rate.
5M+ created · 90%+ finish · After iteration: starts 35% → 70% · +20% trial starts · +90% tier upgrades
Learner knowledge graph 2024–25
Built a Knowledge Graph to track mistakes per learner across core English proficiencies (vocab, grammar, pronunciation) over time. Backfilled proficiency data for 100k+ users, generated proficiency metrics for user profiles, and built lessons automatically based upon recurring mistakes and gaps in knowledge. Fine-tuned several small LLMs to handle high throughput across core error identification pipelines.
+50% trial conversion for personalized-lesson users · up to 30K RPM / 20M TPM
