Andrew Liu

Toolqit is an AI platform that helps IT support companies resolve tickets faster. Owned Toolqit Phone end to end, an AI agent that answers support calls, opens a ticket from the conversation, and connects urgent callers to a technician.

Designed and shipped the long-term memory system for Persona's wearable AI assistant in TypeScript and PostgreSQL. The assistant writes each user's memory as linked markdown files, searched with both keyword and semantic search, so it recalls 2.4x more about each user and finds the right fact 92% of the time, up from 73%.

Built a full-stack tool for Okta that automates SOX compliance audits across Jira, ServiceNow, and Workday. A hybrid pipeline filters tickets with regex, then confirms matches with LLM semantic matching through LiteLLM, cutting manual review time about 90% and saving an estimated 250+ hours a year.

Built and deployed the speech recognition service that scores students reading aloud for a Gates Foundation backed literacy platform that has assessed 9M+ students. It transcribes each recording and lines it up word by word against the passage with edit distance to count words read correctly, running batched inference on a C++ runtime on Google Kubernetes Engine, autoscaled on CPU for school-hour peaks.

Describe a form in natural language and an AI agent builds it instantly. The agent parses intent, generates form schema, and renders a live form in seconds. Every submission runs through a processing pipeline that generates summaries, extracts structured insights, and triggers follow up actions automatically. Built a multi-tenant backend handling concurrent agent sessions with per-user isolation and real time streaming. Got #2 Product of the Day on Product Hunt with 18,000+ users.

Operations platform managing 3,000+ users across 120+ locations. Tracks session attendance, logs volunteer hours, and generates real time reports for every site. 1,200+ weekly sessions and 60,000+ hours logged annually.

AI agent that pulls WarnMe alerts and live incident reports, writes a human readable update, and plots every incident on a map refreshed every 5 minutes. Built after receiving a WarnMe email with no indication of where it actually happened.

Syncs classes from bCourses, Gradescope, and Pensieve into Google Calendar automatically. Built freshman year to solve a problem every Berkeley student has: assignments scattered across four platforms with no single view.

$2,500 prize, presented by Anthropic, Coinbase, and Solana Foundation. Built Lapis, a platform where AI agents score startups on traction and pitch quality and a prediction market sets the valuation, with equity settled on the XRP Ledger.

$900 prize. AI agents like Meta's Muse and Instinct are already buying for millions of people, but most businesses have no way to sell to them. Built Penny, which lets any business get paid by AI agents over instant stablecoin rails, so customers get 3% off and businesses pay up to 98% less in fees than with Stripe or Square.

$1,000 prize. Built Beagle, an AI agent that makes dynamic insurance accessible to businesses. It researches a business risk, matches it to a Gemini prediction market contract field by field, sizes a hedge against the live order book, and buys it once the owner approves the exact plan.

$500 prize. Won the Sui Foundation track solving onchain CTF challenges and exploiting smart contract vulnerabilities across the Sui object model.

Meta's open source deep learning framework used to train most of the frontier AI models in production today, including at companies like OpenAI, Google DeepMind, and Hugging Face. The majority of ML research and production training pipelines run on PyTorch.

• Traced a precision mismatch in TorchInductor where float scalars were kept at full float32 precision instead of being rounded to the model's lower precision, causing loss functions to evaluate incorrectly at numerical boundaries and corrupt gradients during backprop.
• Fixed torch.compile silently dropping channels-last memory format on CPU for max_unpool2d, restriding the decomposition's output to match eager instead of always returning contiguous strides.

The container runtime that runs the vast majority of production software deployments worldwide. Almost every cloud workload, from startups to Fortune 500s, runs inside Docker containers orchestrated on Kubernetes.

• Fixed rootless port binding failing when a container used distinct loopback addresses on the same port. The rootlesskit client was collapsing all loopback IPs to 127.0.0.1, so the host saw duplicate bindings and rejected the second.
• Removed a redundant container store lookup in Swarm service-binding activation. The container was already resolved earlier in the activation flow but was being fetched again by name on every network join.
• Cleaned up loopback IP handling in the rootlesskit client after the port binding fix landed.

The distributed key-value store that every Kubernetes cluster depends on to store its state, including clusters running on Google GKE, Amazon EKS, and Azure AKS. It's the backbone of the Kubernetes control plane and runs underneath most production cloud infrastructure.

• Fixed unbounded log-message cardinality in request tracing. Per-request identifiers like trace IDs were embedded directly in the message string, making every request produce a unique log line. Moved them into structured fields so the message stays stable and can be aggregated.

Courses: Data Structures and Algorithms, Computer Security, Operating Systems, Database Systems, Machine Learning