Proved that direct coordinate output from autoregressive LLMs is unstable due to spatial lookahead planning failures. Built deterministic wrappers (LLM → Python / Matplotlib / RDKit / Schemdraw) where domain libraries handle layout constraints. Token cost: ₹10–30 → ~₹0.5 per video. Errors and latency dropped dramatically.

Engineered Swiss-tournament-style matchmaking to rank JEE question banks using simulated intransitive preference systems. Comparison overhead: O(N²) round-robin → O(N log N). Core finding: pairwise comparison design matters as much as the ranking algorithm itself.

Designed benchmarking rigs for generation-verification loops using DeepSeek and Qwen. Implemented error categorization separating protocol failures (JSON truncation from token limits) from semantic rule failures (business logic errors) to prevent loop decay.

Evaluated serving performance, memory constraints, and scalability tradeoffs using vLLM (RTX 3090), LoRA fine-tuning, and Mixture-of-Experts routing. Built intuition for when training-time architecture choices constrain serving options.

Embedding-based retrieval using cosine similarity, paired with a two-agent Llama 3B reranker hosted locally on an RTX 3090. Balanced retrieval quality against serving latency under real resource limits.

Designed a custom category-overlap evaluation metric — then mathematically invalidated it after discovering structural boundaries in the embedding space that made the metric misleading. Shipping a metric you later disprove is harder and more valuable than one that just looks good.

End-to-end Stable Diffusion implementation project. Guiding students through: Probability Foundations → VAEs → DDPM → CLIP Conditioning. Building onboarding resources and weekly code check-ins that prioritize mathematical understanding over API calls.