Nikhil B.

I build production AI agents and LLM systems. 50+ projects across fintech, healthcare, media, legal tech, climate tech, and research: multi-agent systems, agent harnesses, evaluation systems for agentic workflows, Claude Code, Claude Agent SDK, OpenAI Codex, Hermes, Openclaw, MCP servers, RAG at scale, fine-tuning, and document AI.

7+ years shipping ML and AI in production. $700K+ delivered. Expert-Vetted (top 1%). 100% Job Success Score.

I’m not new to ML. Before the LLM wave I spent years building classical ML and NLP systems in Python: search engines, recommendation engines, entity extraction, knowledge graphs, and predictive models. That foundation means I know when a problem needs an LLM and when it doesn’t. I’ve also cofounded two startups, so I think about business outcomes, not just model accuracy.

Recent work:

• N1 Healthcare: Led the Report Generation team on a medical AI platform. Built the agent harness that dispatches 13+ clinical report workflows across Claude Agent SDK, OpenAI Agents SDK, and Agno through a LiteLLM gateway, with a planner coordinating 20+ specialty agents. Report cost went from roughly $80 to $15-25 per run. Also worked on the upstream medical record parsing pipeline.

• Newsweek (4+ years): 15+ AI applications and 6 Azure Function Apps used daily by non-technical editors. Cosmos DB vector search with Reciprocal Rank Fusion hybrid ranking on 3072-dim embeddings, MCP servers exposing agent tools to Copilot Studio, automated article-to-video pipeline at 95%+ success. A live 30-day window measured 497 editorial hours saved at 15.8% AI adoption. Trained editorial staff to run the systems without developer support.

• Stratifi: Built the AI layer on a Django + Postgres backend. Dual-LLM PDF extractor (GPT-4o + Gemini) at 95%+ accuracy on 50+ page brokerage statements, 10-30 positions/second, with a 24-document eval harness gating prompt changes. Multi-agent financial chatbot routing across 4 market data sources, plus a 4-layer portfolio optimization agent and a Lambda support agent with tool calling. Elasticsearch hybrid search over millions of securities.

• CGIAR: Research impact assessment processing 9,166+ records with 6 fine-tuned GPT-4o-mini variants, GPU-accelerated PDF layout detection with Detectron2, 3-tier evidence extraction at 50 concurrent operations.

• Adoro: 7-stage AI automation pipeline for marketing, turning a week-long manual process into under an hour, processing 100K+ items per run with 3-pass LLM extraction (initial, reflection, self-reflection).