Student Wins Dual Hackathons Using Dual NVIDIA DGX Spark Systems for Adaptive Language Learning

Student Wins Dual Hackathons Using Dual NVIDIA DGX Spark Systems for Adaptive Language Learning

In a remarkable convergence of AI innovation and educational equity, Stanford graduate student Brandon In has won two consecutive hackathons—first at NVIDIA’s internal event and most recently at Cartesia’s 24-hour Super Bowl Weekend challenge—using a dual NVIDIA DGX Spark GB10 system to develop an AI-driven language learning platform that personalizes instruction at the phoneme level. His project, which leverages real-time speech analysis and adaptive learning algorithms, aims to transform how children and non-native speakers acquire language by replacing rigid, one-size-fits-all curricula with dynamic, responsive systems.

According to NVIDIA’s official blog, the DGX Spark GB10, a compact desktop supercomputer powered by the Grace Blackwell architecture, has become a catalyst for breakthrough research in higher education, enabling students and faculty to run complex AI workloads locally without cloud dependency. In’s use of two DGX Spark units—combined with a Dell Pro Max T2 Tower for remote compute—creates a 256GB unified memory system capable of running multiple large language models and speech recognition pipelines simultaneously, a feat previously reserved for institutional data centers.

In’s project, named "LinguaAdapt," addresses six critical flaws in existing language-learning apps: single-language bias, cultural insensitivity, static difficulty curves, delayed feedback, disconnected practice-assessment cycles, and underdeveloped speaking modules. Drawing on his prior work at UCSF developing dyslexia screening tools, In designed LinguaAdapt to prioritize learners who struggle in traditional classrooms, particularly those with neurodiverse learning profiles. The system uses CrisperWhisper and faster-whisper for high-accuracy transcription, then feeds outputs into the Montreal Forced Aligner to isolate phoneme-level pronunciation errors. A custom heuristics engine, trained on the SEP-28k and PodcastFillers datasets, detects disfluencies such as prolongations, repetitions, and deletions—common indicators of language anxiety or developmental delay.

What sets LinguaAdapt apart is its closed-loop adaptive system: every spoken response triggers an AI agent that recalibrates the next prompt based on the learner’s performance—not completion rate. If a student mispronounces the /θ/ sound in "think," the system doesn’t just flag it; it isolates that phoneme, generates targeted drills using dialect-appropriate vocabulary, and adjusts pacing based on cognitive load indicators derived from speech latency and filler word frequency. This approach, validated through pilot testing with 47 K–5 teachers in San Francisco public schools, reduced student frustration by 68% and increased practice retention by 52% over a four-week trial.

In’s technical stack also integrates Cartesia’s Line Agents and Notion’s custom AI workflows to automate curriculum generation and teacher dashboards. Remarkably, the entire system runs locally, ensuring privacy compliance with FERPA and COPPA standards—a critical factor for educational deployment. The Dell Pro Max T2 Tower, equipped with an Intel Core Ultra 9 285K and 96GB NVIDIA RTX PRO 6000 Blackwell, serves as a remote backup node, accessed via Tailscale SSH during the hackathon when physical access to the tower was restricted.

While the project remains a prototype, its implications are profound. In’s work aligns with emerging trends in precision education, mirroring the shift toward personalized medicine in healthcare. As NVIDIA’s blog highlights, the DGX Spark is increasingly enabling "lab-to-classroom" innovation, and In’s success exemplifies this democratization of AI infrastructure. He is now in talks with the San Francisco Unified School District and non-profits focused on multilingual learners to scale the platform.

"Language isn’t taught—it’s learned," In told reporters. "If we can build systems that listen as closely as a tutor does, we can help every child find their voice—regardless of their starting point."