Immanuel Trummer

Database Prof at Cornell. I make data analysis more efficient and more user-friendly.

[Immanuel Trummer @ImmanuelTrummer]

🚗ThalamusDB counting pictures of red cars in the database:
- Semantic operators are described in natural language and evaluated via GPT-5
- Simply store paths to images or audio files in your database – ThalamusDB recognizes the file format and selects the right LLM
💾 Code: https://t.co/6l1gS8zZFA
📄 Website: https://t.co/33PAzsmI7T
#SemanticQueries #ApproximateProcessing #LLM #GPT5 #ThalamusDB

[Immanuel Trummer @ImmanuelTrummer]

✈️ Looking forward to an exciting #SIGMOD2026!
👇 Schedule below.
🇮🇳 See you soon in #Bengaluru!
#Databases #LLMs #LanguageModels #DatabaseBenchmarking #QueryOptimization #QuantumComputing
@lojil192574 @sigmod @SIGMODConf

[Immanuel Trummer @ImmanuelTrummer]

💡 Two arXiv papers published in recent days (one from us, one from TUD) reach the same conclusion: LLMs can now generate C++ code for SQL processing that outperforms classical database systems.
⚙️ Our code generator is based on Claude Code and exploits multiple agents working in parallel. Each agent performs tasks typically associated with different components in a #DBMS, such as workload analysis, query optimization, or physical design tuning.
📊 We compare to various classical #DBMS such as DuckDB, ClickHouse, Umbra, MonetDB, and PostgreSQL, finding that the agent-generated code is often significantly faster. Code generation costs are moderate (<$20), making the approach practical for frequently executed queries.
🤖 Analyzing generated code, we find that agents exploit various optimization techniques, including query-specific data structures, as well as low-level optimizations that are specific to the hardware cache hierarchy of our server.
📃 Paper: https://t.co/PBLC9XX5Tv
💾 Code: https://t.co/MvdH8lHT6F
🌐 Site: https://t.co/2WM2gedwbi
@lojil192574 #LLM #Databases #AI #DB

[Immanuel Trummer @ImmanuelTrummer]

@adwiteekk @freeCodeCamp Glad to hear it 😀

[Immanuel Trummer @ImmanuelTrummer]

Happy to hear you liked it, @DanKornas!

[Immanuel Trummer @ImmanuelTrummer]

A demo of #ThalamusDB (#SIGMOD2023), introducing semantic filter operators. Users write SQL queries with natural language predicates on table columns containing 🖼️ images, 📃 text, or 🔊 sound files. These predicates are evaluated via #LLMs.
In the video (below), I'm querying for furniture ads with pictures showing "wooden tables". After entering my query, #ThalamusDB
1️⃣ performs data profiling and cost-based optimization,
2️⃣ shows the Pareto frontier of cost-quality tradeoffs,
3️⃣ updates bounds on query aggregates while processing.
#ThalamusDB is designed from the ground up for approximate processing, prioritizing data that maximally reduces approximation error per cost unit.
🪧 #SIGMOD2023 demo: https://t.co/hELlBtKRZb
📃 #SIGMOD2024 paper: https://t.co/jzfd4IUbGt
💾 Code repository: https://t.co/zhRn5xX4Km
@SaehanJo @sigmod #GPT4 #LanguageModel #MultimodalData @Cornell @CornellCIS

[Immanuel Trummer @ImmanuelTrummer]

📢 All our posters & talks at #SIGMOD2025!
1️⃣ λ-Tune — using #LLMs to write configuration scripts for databases.
🪧 Poster: https://t.co/plBOp9tSsK
💬 Slides: https://t.co/LiXHMlzNBS
@giannakourisv
2️⃣ SpareLLM — selecting #LLMs with optimal cost-quality tradeoffs
🪧 Poster: https://t.co/7ssHX4fhNU
@SaehanJo
3️⃣ SQLBarber — generating custom benchmarks via #LLMs
🪧 Poster: https://t.co/B3R3wOVX18
@lojil192574
4️⃣ CEDAR — cost-efficient data-driven claim verification via #LLMs
🪧 Poster: https://t.co/NqfPw8Y6CH
@Tharushi96
5️⃣ SwellDB — generating data on-the-fly during query processing by #LLMs
🪧 Poster: https://t.co/bYbByTEzZQ
@giannakourisv
6️⃣ Query optimization for hybrid classical-quantum workflows
💬 Slides: https://t.co/h2ydJZb8zm
7️⃣ Quantum annealing for optimal data partitioning
💬 Slides: https://t.co/OPMI0e0sog
8️⃣ Panel "AI for Future Databases" with @tim_kraska, @adityagp, @feifei_initiald, @ailamaki, and #SurajitChaudhuri
💬 Slides: https://t.co/WvMwknlJLS
@SIGMODConf @sigmod @Cornell @CornellCIS

[Immanuel Trummer @ImmanuelTrummer]

Really proud of my students — @SaehanJo, @giannakourisv, @lojil192574, and @Tharushi96 (left to right) — who each presented their latest work at #SIGMOD2025. Many thanks to the organizers for an amazing conference! @SIGMODConf @sigmod

[Immanuel Trummer @ImmanuelTrummer]

🥳Looking forward to an amazing #SIGMOD2025 conference! Our schedule:
📃 Sunday, 15:00-17:30: Data partitioning with quantum and digital annealers
📃 Sunday, 15:00-17:30: Optimizing hybrid quantum-classical processing pipelines
📃 Tuesday, 10:30-11:30: SpareLLM - selecting LLMs with optimal cost-quality tradeoffs
🖥️ Tuesday, 11:30-13:00: Demonstrating SQLBarber - generating custom benchmarks via LLMs
🖥️ Tuesday, 11:30-13:00: Demonstrating SwellDB - generating data on-the-fly during query processing
📢 Tuesday, 16:30-18:00: Panel on AI for future databases with @TimKraska, @drfeifei, @adityagp, @ailamaki, and Surajit Chaudhuri
📃 Thursday, 10:30-11:30 & 16:30-18:00: λ-Tune - using LLMs to write configuration scripts for databases
🖥️ Thursday, 16:30-18:00: Demonstrating CEDAR - cost-efficient data-driven claim verification
@giannakourisv @Tharushi96 @SaehanJo @lojil192574 @SIGMODConf @sigmod #LLM #SQL #Database

[Immanuel Trummer @ImmanuelTrummer]

💵Don't overpay when using #LLMs! Introducing our upcoming #SIGMOD2025 paper on #SpareLLM by @SaehanJo ...
@sigmod @SIGMODConf #LanguageModel #GPT4 #ChatGPT #CostOptimization #Data @Cornell

[Immanuel Trummer @ImmanuelTrummer]

Outperforming various baselines, including #QuantumAnnealers and classical optimization, for large problem instances with 1000 queries.
#SQL #DB #Quantum #QueryOptimization

[Immanuel Trummer @ImmanuelTrummer]

🥳Paper accepted at #SIGMOD2026! Our paper leverages #DigitalAnnealers (hardware accelerators for optimization) for #QueryOptimization. We scale up to large problem instances using 1⃣domain-specific problem decomposition and 2⃣pre/post-processing on classical machines. @sigmod

[Immanuel Trummer @ImmanuelTrummer]

Had a great time at the @dagstuhl seminar on Table Representation Learning! Lots of interesting discussions and future work directions. Many thanks to the organizers (@FrankRHutter @cbinnig @MadelonHulsebos @eisenjulian)! #TabularFoundationModel #AI #ML #DB

[Ibrahim Sabek @ibrahim_sabek]

The submission deadline for the Q-Data Workshop has been extended by one week. The new submission deadline is April 27, 2025. @SIGMODConf #SIGMOD2025

[Immanuel Trummer @ImmanuelTrummer]

@shctechnologies @ManningBooks @OpenAI @langchain @llama_index Thank you!

[Immanuel Trummer @ImmanuelTrummer]

🥳I finished my book!
📘"Data Analysis with LLMs" shows how to analyze (📄text/🖼️image/🔊audio/📽️video/...) data with #LLMs and #Python!
🔗https://t.co/6tRALMGsGo
@ManningBooks #LLM #GPT4 @OpenAI #SQL #GraphData #AgenticAI @langchain @llama_index #Multimodal #DataScience

[Immanuel Trummer @ImmanuelTrummer]

The book is a hands-on introduction to #LLMs and #Multimodal #DataAnalysis, based on a few mini-projects. It covers the @OpenAI #Python library, #Prompting, #FewShotLearning, #FineTuning, #LLM #Agents, and recent #LLM frameworks like #LangChain and #LlamaIndex.

[Immanuel Trummer @ImmanuelTrummer]

Well deserved 😀

[Immanuel Trummer @ImmanuelTrummer]

🥳Many congrats to Dr. Saehan Jo!
🎓Saehan successfully defended his PhD thesis "Efficient Data Systems for Scalable Analysis with LLMs", introducing systems like #ThalamusDB and #SpareLLM that scale up processing with #LLMs to very large data sets!
@SIGMODConf #Data #SQL #ML