Chaitanya Anand
chaitanyaanand.p24@iiits.in // AI & HCI Researcher // IIIT Sri City
Recruiter Summary: CS undergrad at IIIT Sri City and researcher at the HCI Lab under Prof. Himangshu Sarma, seeking roles in AI/ML engineering and AI/HCI research. Authored DPLA-Retriever (budgeted multi-hop RAG with one Llama-3-8B reader call on a single RTX 4090: 54.2 F1 on HotpotQA with 664ms retrieval; under review at IP&M, Elsevier), built and deployed xoult (live pgvector paper recommender). Co-authored 3 accepted peer-reviewed papers across IEEE ISMAR 2026 (XR Badminton), ACM CHI EA 2026, and IEEE ISMAR 2025. Experienced across PyTorch, CUDA, ModernBERT, FAISS, and real-time HCI architectures.
Education
Indian Institute of Information Technology (IIIT), Sri City
2024 – 2028Bachelor of Technology (B.Tech) in Computer Science & Engineering
- Key Coursework: Data Structures, Algorithms, Database Management Systems, Computer Networks, Linear Algebra.
- Research focus: Immersive systems, unsupervised rehabilitation platforms, and neural retrieval architectures.
Research & Engineering Experience
Systems & HCI Lab Research Intern
Mar 2025 – PresentIIIT Sri City — Chittoor, Andhra Pradesh
- Spearheaded multi-disciplinary research across machine learning, memory-efficient RAG and neural retrieval, algorithmic compliance auditing, neuro-symbolic AI coaching, and spatial computing under Prof. Himangshu Sarma.
- Engineered XR Badminton (Accepted at IEEE ISMAR 2026): standalone Quest 2/3 coaching system that segments overhead strokes into seven phases with a Holt-Linear filtered state machine, scores ten expert-derived rules, simulates quadratic-drag shuttlecock physics, and routes symbolic diagnoses to Gemini 2.5 Flash for spoken coaching with an offline fallback; N=10 pilot usability study. Ongoing project now extending to Sony Mocopi full-body tracking and predictive haptic cues.
- Authored One Reader Call Is Enough (DPLR) (Under Review, Information Processing & Management (IP&M), Elsevier): architected a budgeted multi-hop retrieval cascade executing exactly one call to an untrained Llama-3-8B reader on a single 24 GB consumer RTX 4090 GPU, delivering 54.2 F1 on HotpotQA and 38.1 EM on NQ while cutting retrieval latency from 1,820ms to 664ms.
- Designed dynamic tolerance adaptation algorithms and joint-level kinematics analysis for unsupervised upper-limb XR rehabilitation; led a 21-participant user study (healthy adults) across 58 sessions. Co-authored accepted research at CHI EA 2026 and IEEE ISMAR 2025.
Selected Projects
Python // FastAPI // PostgreSQL // pgvector // Docker // Langfuse // Open Source
Architected a daily personalized paper recommendation service ingesting ~1,900 arXiv preprints/day across 6 ML categories; embedded via local ONNX (all-MiniLM-L6-v2) and indexed in PostgreSQL with pgvector. Features dynamic Rocchio profile updates, transparent historical citations, and CiteULike-a benchmarked CI regression gates (0.1005 NDCG@10, 0.2728 Recall@50).
DPLA-Retriever (Budgeted Multi-Hop RAG Cascade)
PyTorch // FAISS // Cross-Encoder // Llama-3 // IP&M Elsevier Under Review
Architected a budgeted 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, eliminating multi-turn generator overhead. Combines hybrid dense + BM25 retrieval over 21M Wikipedia passages with cross-encoder re-ranking and a learned gating sufficiency classifier (54.2 F1 HotpotQA, 38.1 EM NQ, 664ms retrieval latency).
Briefed (`briefed-cli` on npm)
TypeScript // Node.js // Git Hooks // Vitest // Open Source
Developer CLI tool closing the AI context gap on `git pull`. Coordinates Git hook triggers with atomic file locking and non-blocking incremental workspace indexing for agentic editors like Cursor and Claude Code, backed by an 89-test Vitest suite.
Minecraft-Helper (System-1 Autonomous Companion Agent)
Node.js // Mineflayer // Jev & Decider (System-1) // TypeSafe AI Architecture // PyTorch // SSE // Open Source
Architected an autonomous game companion replacing multi-second cloud LLMs with a Jev-style System-1 calibrated decision model (TypeSafe AI
/v1/systemone wire protocol and local 2B classifier, ~185ms GPU forward pass) operating over verified code affordances (System-1 intuition + System-2 deterministic planning). Implemented full recipe-tree synthesis, 3D A* navigation, combat kinematics with shield-timing, death retrieval across 97 vanilla death messages, bilingual English/Hinglish intent classification (88% accuracy across 175 evaluation queries), and validated by 198 unit tests.
XR Badminton (Biomechanical Athletic Coaching Platform)
Unity C# // Python // Sony Mocopi 12-IMU // IEEE ISMAR 2026 Accepted
Ongoing systems research project with 1 accepted paper at IEEE ISMAR 2026. Standalone Quest athletic simulator featuring real-time 12-IMU Sony Mocopi forward kinematics streaming at 72.5 Hz, quadratic aerodynamic shuttlecock drag modeling, few-shot predictive haptic cues fired ~200ms before racket contact, and Gemini 2.5 Flash coaching verbalization.
GigsForGigs (Freelance Marketplace Backend)
NestJS // TypeScript // REST APIs // RBAC // Custom Relational Mapping
Designed a modular NestJS freelance marketplace backend featuring strict DTO validation, Role-Based Access Control, and a custom constraint-safe in-memory relational database mapping layer built using JS Maps.
Technical Skills
Languages
ML, RecSys & Search
Backend & Cloud
Core Domains