Chaitanya Anand

Personalized daily research paper recommender ingesting ~1,900 arXiv preprints/day across 6 ML categories. Implements a two-stage retrieval pipeline using local ONNX embeddings (all-MiniLM-L6-v2) and PostgreSQL (pgvector) with dynamic Rocchio preference centroid updates. Features transparent citations (“because you saved X”), full telemetry tracking, and an offline evaluation benchmark on CiteULike-a (5,551 real user libraries) delivering 0.1005 NDCG@10 and 0.2728 Recall@50 with automated CI regression gates.

A standalone Quest coaching platform combining deterministic biomechanical analysis with generative Large Language Models. Segments overhead strokes into a seven-phase sequence using a Holt-Linear filtered state machine and evaluates ten expert-derived rules against procedurally generated templates. Features quadratic-drag shuttlecock physics, a neuro-symbolic feedback engine that routes rule-based diagnoses to an LLM verbalizer (Gemini 2.5 Flash) with an offline audio fallback, challenge-point pacing of feedback modalities, and 3D volumetric replays. Evaluated in an N=10 pilot usability study.

Architects an end-to-end multi-hop retrieval cascade executing exactly one call to an untrained Llama-3-8B-Instruct reader on a single 24 GB consumer RTX 4090 GPU. Combines hybrid dense and BM25 retrieval over the 21M-passage Wikipedia corpus with cross-encoder re-ranking and a learned gating sufficiency classifier. Delivers 54.2 F1 on HotpotQA (above every published FlashRAG method that does not train its generator) and 38.1 EM on Natural Questions, while scoring 205 instead of 368 cross-encoder pairs per question and cutting retrieval latency from 1,820ms to 664ms.

An ongoing VR-based athletic coaching platform and simulator. Tracks swing trajectory using Meta Quest controllers and 12-IMU Sony Mocopi forward kinematics streaming at 72.5 Hz, modeling quadratic aerodynamic shuttlecock drag physics and few-shot predictive haptic cues that fire ~200ms before racket contact when a flawed stroke is forecast. Implements arpeggiated visual overlays for shot quality analysis and dynamic height scaling.

Designing budgeted multi-hop neural retrieval cascades for edge and consumer GPU workstations. Utilizes hybrid dense + BM25 retrieval over 21M passages with cross-encoder re-ranking and a learned sufficiency gate to execute multi-hop RAG with exactly one reader call on a single RTX 4090 (54.2 F1 on HotpotQA, 664ms retrieval latency).

An autonomous companion agent in Minecraft implementing dual-process cognitive architecture (System-1 / System-2). Discards non-deterministic, high-latency generative LLMs in favor of a Jev-style System-1 calibrated decision model (TypeSafe AI /v1/systemone architecture, executing local 2B classifier inference in ~185ms on laptop GPU) operating over code-verified game affordances. Integrates deterministic recipe-tree synthesis, A* 3D spatial navigation, reactive shield-timed combat, proactive buddy-mode assistance, bilingual English/Hinglish intent classification (88% accuracy across 175 evaluation queries), and validated by 198 passing unit tests.