Case studies

Selected work

Production systems, platforms, and products. Each one shipped.

Agentic AI Platform

Microsoft · Bing Whole Page Answer

Problem: the search data I needed (entities, intents, multimodal content) wasn't available from existing sources, and no off-the-shelf agent platform fit the production bar. Approach: built a distributed agent platform — agent orchestration, quality gates / evaluation, multimodal extraction, and a browser-agentic crawler — with ~25 production agents across markets, reusing shared components per vertical. Result: a platform that lets us add new verticals (and new extractors) without rebuilding, with measured quality before anything reaches users.

  • Offline entity discovery + intent mining + multimodal data extraction.
  • Quality gates, evaluation, and reusable agent templates.
  • Why hand-rolled: production bar, data diversity, and vendor lock-in — not because no framework existed.
AgentsLLMRAG CrawlingEvaluation

Copilot Studio / Power Platform & Governance

Microsoft ecosystem · Enterprises

Problem: enterprises want AI agents but are blocked by security and governance — who can act, what's audited, how do you approve risky steps? Approach: Copilot Studio / Power Platform agents wired into Azure Functions and enterprise data, with Entra ID RBAC, Dataverse row-level security, and human-approval gates before any consequential action. Result: agents that are safe to put in front of real users — with audit trails, not just a demo.

  • Copilot Studio + Power Automate / Apps + Azure Functions + C#.
  • Entra ID auth, identity, RBAC, row-level security, compliance/audit.
  • Connecting agents to SAP / CRM / SharePoint with approval gates.
Copilot StudioPower Platform Entra IDAzure Functions

News Feed Recommendation Platform

Microsoft · MSN (Edge new-tab / Win11 feed)

Problem: a recommendation feed that's personalized, fast, and well-laid-out across user and location signals. Approach: an orchestrator service that layers data — user/location, article IDs from the recommendation API (plus static and per-user data behind a cache), vertical content (weather, sports, finance, travel, shopping, real estate), and topics. Two-stage ranking: content ranking for articles, then whole-page optimization (layout / placement) for what should be most prominent.

  • Orchestrator + cache layer over article static & personalized data.
  • Content ranking (article) and layout ranking (whole-page) stages.
  • High concurrency, low latency, graceful degradation.
RankingRecommendation CacheDistributed Systems

Real-Time Ad Platform & Data Platform

Hulu (Beijing / US) · Streaming ads

Problem: streaming ads need real-time decisioning and targeting while the data platform feeds and verifies it. Approach: built on the ad decisioning / serving path — targeting, ranking, and budget rules at millisecond scale — with the ad data platform (ETL, feature/sample pipelines, model training/inference) feeding it. Result: a connected chain from data platform → decision → serving that shares a single source of truth for ads.

  • Real-time ad decisioning & serving, targeting, ranking.
  • Ad data platform: features, samples, model train/serve.
  • High throughput, low latency, correctness under load.
Ad TechReal-time RankingData Platform

Interview Assist — AI-native desktop product

Self-built · macOS (Intel + Apple Silicon)

Problem: I wanted an AI that captures a question (audio/screen/clipboard), turns it into text, and returns a structured answer in seconds. Approach: a two-machine macOS product — a sensor on the interview laptop (system-audio/screen capture, OCR, clipboard) and a display on a second machine, connected over LAN, with pluggable LLM/vision providers, templates per interview type (coding / OOD / system design / behavior), and answer-language & effort controls. Result: a real working product across M1 (Intel) and M4 (Apple Silicon).

  • Audio transcription + OCR + clipboard capture → LLM → structured answer.
  • Pluggable Qwen / DeepSeek vision and text providers; language & effort toggles.
  • Cross-machine LAN sync, session persistence, interview-type templates.
ProductmacOSASR LLMVision

Agent Research & Open-Source Scaffold

Deep research + OSS

Problem: the agent ecosystem moves fast — which harness, platform, and patterns are worth adopting? Approach: deep-dive research into Pi agent (DeepSeek integration), deepseek-harness (agent runway/test harness), phone-harness (phone-use agents), and open Codex. Extracted production patterns: quality gates, model-agnostic providers, approval gates, and reusable tool scaffolds. Result: an open-source, model-agnostic agent harness you can start from (see repo), plus informed vendor choices.

  • Pi / DeepSeek harness / phone-harness / Codex harness deep dives.
  • Open agent scaffold with eval + human-approval gate.
  • Multimodal + multi-model provider abstraction.
ResearchOpen Source AgentsEvaluation