Thomas Li

Own end-to-end technical strategy, architecture standards, engineering, and algorithms for the organization responsible for AI across Microsoft Edge; lead ~20 engineers across the US and China over four product tracks, spanning model training and research. Drove the architecture consolidation of Copilot across Microsoft — unifying Edge and Bing across Windows, Mac, and mobile onto a shared agent architecture and a common prompt-tooling layer.

Product and business delivery

• Tabs Auto Grouping (shipped 2022, ~12 months ahead of Google): the industry's first LLM-based browser tab auto-grouping, ultimately converged onto an in-house SLM (fine-tuned on Mistral) in production. Drove 3× growth on the Tabs business line.
• Contextual RAG: long-document understanding in Edge Copilot.
• End-to-end rebuild of Edge Copilot: unified support for real-time interaction and browser operation, and drove the experimentation and iteration process for Copilot tooling and prompts across Microsoft. Task success 76% → 90%; action success 50% → 90%; first token 1s → 241ms.
• Journeys (shipped 2026): the industry's first browser-native experience turning long-term browsing history into visual task cards, each with suggested next actions, surfaced on the browser home page and continuable inside Copilot; +3.9% conversion, +4.3% CTR, +24% coverage.
• Office Agent / Digital Labor: built a multi-agent stack from scratch covering Office workflows, then extended it into a team-first agent harness combining team-level memory with proactive participation in long-horizon tasks, shipped into Microsoft Teams.
• Web Remix: lets the browser generate web capabilities on demand — mini-apps produced from the current page or browsing history, and natural-language reshaping of pages.
• Model routing and training: own model-routing strategy across accuracy / latency / cost, moving production workloads between frontier models and in-house fine-tuned small models.
• Scale: Edge Copilot grown from 0 to millions of DAU; Journeys and Tabs each operating at million-DAU scale.

Personalization & growth

• Deep Embedding: unified user and content representations. Multi-task learning (MTL) to jointly model multiple objectives, plus hard-negative mining; distillation and quantization compressed the model 200M → 10M.
• Unified personalization and recommendation system: on-device and cloud computation of user intent and representations, serving Notification and Copilot Suggestion Chips — Notification CTR +110%, Chips adoption +120%.
• Growth and paid acquisition: built the User Segmentation framework and owned Edge Mobile growth, DAU 3M → 20M; ARPU +90%, LT_7 +20%, retention +1.4pp, +700K incremental DAU; ROI 291% in GB and 188% in US.

• Live-stream recommendation: DSSM softmax-in-batch, bitwise FM/AutoInt, CGC, PCGrad, GradNorm, ESMM. Follow penetration +17%, watch time +1.15%; doubled the number of active streamers.
• User / family recommendation: bipartite-graph GraphSAGE with GNN, attention, and self-supervised learning. Family-business D+1 retention +3.4%, follow penetration +100%.
• Platform & infrastructure: built the engineering foundation for social recommendation from scratch — graph data-mining platform, feature platform, retrieval platform, and the data reporting system; unified retrieval, feature, bandit, and traffic-control services across 15 entry points and 4–5 verticals; overall penetration +10%.

• Musician Recommendation: built from scratch (word2vec similar-user retrieval + LR ranking); follow rate and follow count both +100%.
• Follow-feed ranking: rebuilt with itemCF + GBDT; play count +10%.

• Tianhe Program: modeled the business constraints as an online optimization problem, guaranteeing delivery volume while holding CTR and display revenue flat.
• Taobao Push: end-to-end ownership across data analysis, online delivery engineering, and the operations management platform. Solved constrained delivery allocation via Lagrangian optimization — lifting CTR while guaranteeing each business line its committed traffic share; +100% over the prior audience-targeting strategy.
• Coupon distribution: audience targeting and allocation of shopping coupons under budget and redemption constraints.
• Double 11 Everest Program: owned demand forecasting and traffic control. Forecasting drove sell-through completion above 90% for every merchant in every time window across Double 11.

• Data collection: built the pipeline forwarding on-premise data into an HDFS cluster; 10× throughput improvement.
• Web: web development for early Huawei Cloud.
• C++ backend: design and development of C++ network-element management software.