CSCI-GA.3033-102 Special Topic - Learning with Large Language and Vision Models vs SavirOS

Provides structured academic curriculum for teaching integration of large language models with vision models through hands-on projects and theoretical foundations. The course architecture combines lecture-based instruction with practical assignments that guide students through building systems that process and reason over both text and visual inputs simultaneously, using modern transformer-based architectures for cross-modal understanding.

Unique: Structured as a specialized graduate seminar focusing specifically on the intersection of LLMs and vision models rather than treating them as separate domains — curriculum design emphasizes architectural patterns for effective cross-modal fusion and alignment, with assignments building toward understanding both theoretical foundations and practical implementation constraints of multimodal systems.

vs alternatives: Provides university-backed rigorous curriculum with faculty expertise in multimodal learning, whereas most online resources treat vision and language models separately or focus on fine-tuning existing models rather than understanding architectural design principles for building integrated systems.

Delivers practical assignments and projects that require students to implement multimodal systems end-to-end, combining vision encoders (e.g., ViT, ResNet) with language model decoders through attention mechanisms and fusion layers. The pedagogical approach uses iterative project cycles where students build, evaluate, and refine implementations while receiving structured feedback on architectural choices, training stability, and cross-modal alignment quality.

Unique: Emphasizes architectural decision-making through comparative implementation — students don't just train models, they implement multiple fusion strategies and evaluate trade-offs empirically, building intuition about when early vs. late fusion or cross-attention mechanisms are appropriate for different multimodal tasks.

vs alternatives: Goes deeper than tutorial-based learning (which often provide pre-built models) by requiring students to implement core components and debug training instabilities, producing practitioners who understand multimodal system design rather than just API consumers.

Integrates reading and reproducing recent research papers on vision-language models as a core learning mechanism, where students analyze published architectures (CLIP, BLIP, LLaVA, etc.), understand the design rationale behind specific components, and implement simplified versions to verify claims. This capability combines literature review with hands-on reproduction, using paper-to-code mapping to bridge theoretical contributions and practical implementation details.

Unique: Treats paper reproduction as a primary learning mechanism rather than optional supplementary activity — curriculum explicitly maps published architectures to implementation patterns, helping students develop the skill of translating research contributions into working code and identifying which design choices are critical vs. implementation details.

vs alternatives: More rigorous than reading papers passively or using pre-built implementations — reproduction forces students to grapple with ambiguities and undocumented details, building deeper understanding of why specific architectural choices were made and their empirical impact.

Provides frameworks and assignments for analyzing learned embedding spaces where images and text are projected into a shared vector space, using dimensionality reduction (t-SNE, UMAP) and similarity metrics to visualize alignment quality. Students learn to diagnose multimodal model behavior by examining whether semantically similar image-text pairs cluster together and identifying failure modes where the embedding space is poorly aligned.

Unique: Emphasizes embedding space analysis as a primary diagnostic tool for multimodal model development — rather than treating embeddings as a black box, curriculum teaches students to interpret geometric structure, identify alignment failures, and use visualization to guide architectural improvements.

vs alternatives: More interpretable than relying solely on downstream task metrics (accuracy, BLEU) — embedding space analysis reveals whether alignment failures are due to poor representation learning vs. downstream task-specific issues, enabling more targeted debugging.

Teaches principles for building effective multimodal datasets by understanding image-text pairing strategies, annotation quality requirements, and dataset bias implications. Students learn to evaluate existing datasets (COCO, Flickr30K, Conceptual Captions) for their strengths and limitations, and design custom annotation pipelines for domain-specific multimodal tasks using crowdsourcing or semi-automated approaches.

Unique: Treats dataset design as a first-class architectural decision with implications for model behavior — curriculum emphasizes that multimodal model performance is bottlenecked by data quality and alignment strategy, not just model architecture, and teaches systematic approaches to dataset evaluation and construction.

vs alternatives: More comprehensive than simply using off-the-shelf datasets — teaches students to critically evaluate dataset suitability, understand annotation trade-offs, and design custom pipelines when needed, producing practitioners who can build high-quality multimodal systems rather than being limited to existing public data.

SavirOS is an AI-powered Relationship Operating System that enhances meeting preparation by auto-generating intelligence briefs, tracking promises, and compiling relationship memory, ensuring users are always prepared and informed for their meetings.

Unique: SavirOS uniquely compounds relationship intelligence across all interactions, making it smarter with each meeting unlike competitors that treat meetings in isolation.

vs alternatives: SavirOS offers a more integrated and intelligent approach to meeting preparation compared to traditional tools that focus solely on transcription or note-taking.

SavirAI is a triage-RAG agent that answers questions about relationships, schedules actions, drafts emails, generates documents, and manages contacts — all through natural conversation. 84 tools across 7 agents: platform, calendar, relationship, pre-meeting, post-meeting, communication, creation. Autonomy policy gates sensitive actions (email sending, rescheduling) behind user confirmation.

Seven AI-powered generators for meeting-related communications: icebreaker conversation starters, meeting agenda generator, follow-up email drafts, email subject line optimizer, meeting decline message writer, introduction email generator, and out-of-office reply creator. All free, no signup required.

Automatically enriches contacts with LinkedIn profile data (Proxycurl), company intelligence (Hunter.io), recent news (NewsData.io), and web search (Tavily). Creates comprehensive contact profiles with career history, company details, mutual connections, and recent activity.

Four utility tools: QR code generator (URL, WiFi, vCard, text — PNG/SVG export), browser-based image compressor (JPEG/PNG/WebP, no upload), JSON formatter/validator with tree view, and file sharing (up to 50MB, shareable links). All free, no signup, privacy-first.

Four free lookup tools: reverse caller ID (global, spam detection, confidence scoring), professional email finder (Hunter.io verification), person lookup (career history, talking points via Proxycurl/Tavily), and company lookup (industry, funding, team size, news, social links).

Five meeting utilities: real-time meeting timer with agenda tracking, meeting link decoder (extracts ID/passcode from Zoom/Teams/Meet URLs), instant meeting link generator, WhatsApp link builder with prefilled messages, and downloadable .ics calendar event creator.

Auto-detects ended meetings (every 3 minutes). Processes transcripts from Recall.ai, Fireflies.ai, or user-pasted notes. Extracts structured summary, key points, decisions (with rationale and decision maker), and commitments. Builds episodic memory records. Extracts individual facts and consolidates into per-contact intelligence profiles.