GenAI Systems Lab

They have 2 million legal documents across PDF, DOCX, scanned images, and emails. No two formats are alike. Their current system only handles clean te...

Their RAG system retrieves relevant-looking chunks but the answers are wrong. Why? Critical context is split across chunk boundaries. A drug interacti...

They started with Chroma (in-memory) for their prototype. Now with 50M vectors, it crashes every 6 hours. They need a production vector store that han...

Their semantic search misses exact matches. When developers search for 'CUDA_OUT_OF_MEMORY error code 0x3F', the system returns generic GPU articles i...

Their RAG retrieves 10 chunks for each query but the most relevant one is often at position 5-8, not position 1. Doctors waste time scrolling. For urg...

Their code assistant retrieves 10 relevant code snippets but stuffs them ALL into the prompt. Result: the LLM gets confused by contradictory examples ...

Researchers don't trust AI answers without sources. They need every claim linked to its source document, paragraph, and page number. Current system ge...

Designers search with text ('minimalist blue logo with mountain') but their assets are images. Current system only has filename search. Designers wast...

Customer: 'What's my deductible?' Bot answers correctly. Customer: 'And what about for dental?' Bot retrieves generic dental info, ignoring that the c...

They deployed RAG for 20 enterprise clients but have no way to measure quality. Some clients report 'great answers' while others say 'garbage'. Withou...

Analysts want to ask questions in English and get SQL queries executed automatically. Current approach: LLM generates SQL but it's often wrong because...

Researchers need to answer multi-step research questions: 'Is Drug X more effective than Drug Y for condition Z? What do the latest 3 meta-analyses sa...

PMs want to say 'Set up the new marketing campaign' and have the AI create 15 tasks, assign owners, set dependencies, and create a timeline. ReAct is ...

Their AI assistant forgets everything between sessions. User tells it their preferences on Monday, by Wednesday it asks the same questions again. They...

Creating a blog post requires 4 specialized skills: research, writing, editing, and SEO optimization. One general-purpose agent does all of them poorl...

Data scientists want to say 'Analyze this CSV and create visualizations' and have the AI write AND execute Python code. But unrestricted code executio...

Their trading agent executes trades autonomously. Problem: it once bought $2M of a stock based on a misinterpreted earnings report. They need confiden...

Their DevOps agent automates infrastructure tasks but crashes completely when ANY step fails. A DNS timeout in step 3 of 10 kills the entire workflow....

They deploy 50 different agents for clients. But they have NO WAY to measure if agents are getting better or worse over time. A model update improved ...

They host 200 different agents for 50 enterprise clients. Each agent needs its own tools, memory, and LLM config. They need a platform that manages ag...

They serve Llama-3.1-70B for enterprise clients but throughput is terrible - 5 requests/sec on a $30K A100. Competitors claim 50 req/sec on same hardw...

They use 5 different LLM providers (OpenAI, Anthropic, Google, Cohere, open-source). Each has different rate limits, pricing, and failure modes. They ...

40% of their queries are semantically identical (different wording, same intent). Each still costs a full LLM call. $200K/month wasted on duplicate co...

GPT-4 is great at general tasks but mediocre at their specific legal format. They need outputs in exact legal citation format, jurisdiction-specific l...

Clients ask 'Which model should we use?' and they have no systematic way to answer. They evaluate by vibes - running 10 queries and eyeballing results...

Their medical AI occasionally: 1) Recommends specific drug dosages (illegal without prescription), 2) Leaks patient names from context, 3) Gets jailbr...

AI costs grew from $10K to $180K/month in 6 months. Nobody knows which product, team, or feature is driving the spend. They need real-time cost attrib...

Their clients deploy AI but have no visibility into quality, cost, or performance. When something goes wrong, it takes days to diagnose. They need an ...

They want to test a new prompt (v2) that marketing claims 'converts 20% better'. But they can't just switch - if v2 is worse, they lose revenue. They ...

Their AI powers real-time fraud detection for a bank. If the AI goes down, fraudulent transactions go through unchecked. They need 99.99% availability...

Fortune 500 client with 2M documents across 50 departments. Each department needs their own AI assistant that answers ONLY from their documents. Must ...

E-commerce company handles 50K support tickets/day. 60% are simple (order status, return policy) but humans handle ALL of them. They need AI that reso...

Building a coding assistant that understands the ENTIRE repository context, not just the current file. Must generate code that fits the existing archi...

500K posts/hour. Current moderation catches only 40% of policy violations. False positive rate is 15% (removing legitimate content). They need AI mode...

Building a Perplexity-style AI search: user asks a question, system searches the web, synthesizes an answer with citations. Must handle 1M queries/day...

500K students, 20 subjects. Each student learns differently - some need more examples, some need challenges, some need patience. A one-size-fits-all A...

Trading desk wants AI that reads news, analyzes earnings, and suggests trades - but with strict risk controls. One bad trade can lose millions. They n...

Media company with 10M assets (images, videos, documents, audio). Users want to search across ALL modalities: 'Find the product photo where someone is...

Building a complete 'operating system' for AI: enterprises register agents, define tools, set budgets, manage versions, and monitor everything from on...

Building a research assistant that can: identify gaps in its own knowledge, design experiments to fill those gaps, execute the experiments, evaluate r...