I build machine learning systems end to end, and I care about the whole span of that sentence — the fine-tuning run, the inference server it lands on, and the infrastructure holding both up. Most of my work is on AI² at Agnisys: an EDA verification intelligence platform that reads hardware specifications and produces verification artifacts a person would otherwise write by hand.
Outside work the same instinct shows up as self-hosted infrastructure and small tools written in Go and Rust, usually because something I wanted did not exist.
Agnisys, Inc.
Owns AI², an EDA verification intelligence platform: multimodal document understanding over PDF and Docx specifications, FSM extraction, SystemRDL/XRSL register-map reasoning, and RAG over design documents.
- Fine-tuned and self-hosted a Qwen3.6 deployment, replacing the OpenAI API dependency outright.
- Grammar-constrained generation straight into UVM-ready JSON.
- GA- and RL-driven fine-tuning cut verification turnaround by more than 70%, and HLD-to-driver from 30 days to under 6 hours.
- MCP servers exposing IDS-Verify and IDS-Batch to LLM agents.
IIM
Generative modelling and computer vision. Work shown at RecSys 2023.
- Recommender performance up 25.8% on cold-start, using NLP personality matrices.
- DCGAN-generated synthetic datasets.
- Refining efficiency up 68% via a Multi-Armed Bandit.
JIIT, Noida
- Languages
- Python · Go · Rust · C/C++ · JS/TS · SQL · Lua
- ML
- LLM fine-tuning · Multimodal · RAG · Constrained generation · GA · RL/RLHF · PyTorch · Haystack
- Systems
- Self-hosted agentic workflows · MCP servers · EDA tooling
- Cloud
- AWS (IAM, S3, SQS, SNS, autoscaling) · Oracle Cloud · GCP · Docker · GitHub Actions
- hardikraina079@gmail.com
- GitHub
- github.com/Raina-Hardik
- linkedin.com/in/hardik-raina
- ORCID
- 0009-0007-4281-2989
- Résumé
- Download PDF