Capstone projects built
across our cohorts.

An advanced multi-agent orchestration system designed to automate the high-stakes B2B RFP (Request for Proposal) process. The system utilizes a recursive retrieval strategy (RAG) to ingest thousands of legacy documentation pages and technical specs, synthesizing compliant, context-aware responses in minutes. Features include a verification loop where a 'Reviewer Agent' audits the output for technical accuracy against golden datasets.

A custom-trained 1.5B parameter transformer model specialized for high-throughput Python code generation. Rahul executed the entire pipeline: from writing a custom BPE tokenizer to curating a 10GB 'Fine-Code' dataset. The model was trained using Distributed Data Parallel (DDP) across 4xA100 GPUs and features Flash Attention 2 for efficient long-context window inference during real-time IDE completion tasks.

A production-grade autonomous triage system that manages a 50,000+ monthly ticket volume without human intervention. Built entirely using no-code state machines, it performs semantic classification of incoming Zendesk tickets, extracts high-priority entities (Order IDs, SLA tiers), and routes them to specialized AI sub-agents. The system includes an automated 'Human-in-the-loop' handoff for high-churn risk scenarios.

A specialized RAG engine designed to parse and structure hyper-complex SEC 10-K and 10-Q filings. Unlike standard RAG, this system uses LlamaParse to maintain table hierarchies and accurately extract nested financial metrics. A custom Evaluation (Eval) pipeline was built using G-Eval to ensure that numeric extractions have zero margin for error before being pushed to downstream investment dashboards.

A research environment exploring emergent game-theoretical behaviors in LLM-to-LLM negotiations. Samir implemented two distinct personas—'Skeptical Buyer' and 'Aggressive Seller'—that compete over contract terms within a sandboxed environment. The project tracks 'Tactical Shifts' and calculates 'Utility Gain' for each party, providing a rich visualization of how different prompt-tuning strategies affect multi-turn strategic reasoning.

An air-gapped, privacy-first technical documentation assistant. Designed for enterprise environments where cloud data leakage is a critical risk, this system runs entirely on local hardware using Ollama and ChromaDB. It features a custom 'Context Compression' layer that allows 8B parameter models to reason over massive local knowledge bases without exceeding 16GB of VRAM, maintaining high-fidelity answers offline.