- Senior Machine Learning Engineer at Agnisys Inc., building AI² - an EDA verification intelligence platform spanning multimodal document understanding, register-map reasoning, and RAG over design documentation.
- I train and fine-tune LLMs from scratch and self-host them company-wide - a fine-tuned Qwen3 deployment replaced our OpenAI API dependency for embeddings and synthetic data generation.
- Focused on constrained/grammar-based generation, genetic-algorithm-driven fine-tuning, MCP servers, and domain-specific agentic tooling.
- Manage AWS infrastructure in depth - autoscaling, S3-based backup/DR, and cost optimization.
- Published researcher - 3 papers and a DAC 2026 poster.
- Site: hardikraina.com - projects, publications, and résumé.
- ORCID: 0009-0007-4281-2989 · hardikraina079@gmail.com
Languages:
- Python, Go, Rust, C/C++, JavaScript/TypeScript, SQL, Lua
AI/ML/DL:
- LLM Fine-tuning & Training, Multimodal AI, RAG, Constrained Generation
- Genetic Algorithms, Reinforcement Learning, RLHF, PyTorch, Haystack
Domains:
- Self-Hosted Agentic LLM Workflows, MCP Servers, Generative AI, CV, EDA Tooling
Cloud & Infra:
- AWS (IAM, S3, SQS, SNS, autoscaling), Oracle Cloud, GCP, Docker, GitHub Actions- AI²: EDA Verification Intelligence Platform - Poster, DAC 2026 (submitted to DVCon)
- From Data to Decisions: Predictive Analysis on FAIR Prophet - IC3 2024
- Language Agnostic NMT: Novel Improvements on Transformers - DOI
- Yoga Pose Detection using VGG - Read PDF
- Owning AI² end-to-end - FSM extraction, SystemRDL/XRSL register-map reasoning, and semantic RAG over specs.
- Genetic-algorithm and RL fine-tuning strategies that cut verification turnaround by 70%+ (HLD-to-driver from 30 days to under 6 hours).
- MCP servers exposing the internal EDA toolchain to LLM agents, plus a terminal-native VCD waveform-rendering coding assistant.
- smelt - highly parallel, ffmpeg-powered media transcoding CLI/TUI with hardware-accelerated encoding (Go)
- xml-bifurcate - massively parallel Rust tool that isolates the exact failing block in a chip design; adopted internally at Agnisys (Rust, Tree-sitter)
- Image Super-Resolution Service (EDSR) - production-shaped SISR service (PyTorch, FastAPI, Docker, Caddy)
- mcat contributor - fixed three interactive-mode bugs in a 1.3k-star Rust terminal media viewer (PR #83, merged)
