At Alignerr, I author deterministic, solver-backed RL training environments for LLM post-training. My tasks include structural engineering and aerodynamics design, with grading pipelines designed to resist reward hacking and adversarial submissions.
I built Never Ghost, an AI front desk for solo contractors that qualifies leads and sends summaries for owner approval before booking. Its text-transformation pipeline uses Gemini Pro with strict JSON schema validation.
In my Agentic Bookkeeping System venture, I built an autonomous platform for transaction categorization, anomaly detection, financial forecasting, and financial statement generation. I also implemented role-based access control, data encryption, and ingestion from CSV files and bank feeds.
For Manas, I designed and fine-tuned a 7B mental health companion for Indian users on a custom conversation dataset. I built an RLAIF-style evaluation pipeline and iterated on dataset quality, achieving a 45% auto-approval rate on unseen conversations.
