Amaresh Hebbar

AI/ML engineer with hands-on production experience fine-tuning LLMs, building supervised fine-tuning (SFT) and preference-pair (ORPO) datasets, and shipping model-evaluation and safety pipelines. Trained and published 20+ open-weight models on HuggingFace across medical, retrieval, and code domains. Comfortable with high-volume, detail-critical annotation and QA work across Python, Go, TypeScript, and JavaScript codebases, and with writing/reviewing structured labeling guidelines.

Seeking freelance/contract AI training, RLHF, data labeling, and LLM evaluation roles.

AI Engineering Lead — Fine-Tuning, Evaluation & Dataset Pipelines at Studio Ilios / Humanova Minds (2025-01 – 2026-06)

• Trained and published 20+ open-weight LLMs (Qwen2.5, Llama, Mistral-based) using a QLoRA → DoRA → ORPO → merge pipeline, covering medical coding, retrieval, and reasoning domains.
• Built and manually verified large supervised fine-tuning and preference-pair (ORPO) datasets from raw source documents (CMS regulatory XML, PubMed-derived text, clinical notes) — deduplication, schema validation, and quality review at 400K+ row scale.
• Designed and ran a 13-stage LLM safety and evaluation pipeline including a hallucination-detection firewall (94% catch rate), directly comparable to hallucination-audit and reasoning-verification labeling work.
• Wrote labeling/annotation guidelines and rubrics for internal team members and interns producing model training data, then reviewed their output for consistency and quality (QA on annotator work).
• Reviewed and debugged model-generated code across Python, Go, and TypeScript/React Native as part of production pipelines — equivalent to code-generation review and correctness verification.

B.E. in Computer Science & Engineering – Dayananda Sagar College of Engineering (2021 – 2025)