Sushant Joshi
I build production AI systems that users trust enough to run without me.
What I build
I build agentic AI platforms, RAG pipelines, and LLM evaluation frameworks that ship to production and stay running.
My systems have been adopted by enterprise clients for daily operations without engineering support.
I operate as a sole engineer: architecture, deployment, documentation, handoff.
github.com/SushantTusharJoshiProjects
Sole engineer on a production agentic AI platform at a global enterprise IT services firm. Built on AWS Bedrock with Claude API. Designed end-to-end RAG pipeline with Hugging Face embeddings, FAISS vector indexing, cosine similarity retrieval with confidence thresholding, and rubric-level chunking strategy that outperformed page-level chunking by 25% on relevance metrics. Implemented multi-step agent orchestration via MCP with tool-calling for autonomous data retrieval from external enterprise systems. Built LLM evaluation framework for non-deterministic outputs: tolerance-based regression testing against 50 golden examples, structural constraint validation, and retrieval confidence monitoring with automated deployment gating. Flask REST APIs, Docker, CI/CD via GitHub Actions, 90%+ pytest coverage. Client team adopted the system for daily use without engineering support, improving assessment accuracy by 30% and cutting shortlisting time by 40%.
System architecture
Independently built a HIPAA-compliant clinical risk prediction platform. XGBoost + LightGBM ensemble trained on 200K+ synthetic patients with 5 years of longitudinal adherence data in PostgreSQL. Domain-driven feature engineering with composite risk scoring from comorbidity indices, medication load, and admission history. SHAP explainability so clinicians see exactly why a patient was flagged. Groq-powered AI narratives that translate model outputs into clinical language. KNN similar-patient matching across demographics. Demographic bias auditing across patient groups with disparate impact analysis. Human-in-the-loop audit workflows for regulated clinical decision-making. Caught a data leakage bug where an earlier training run hit AUC=1.0 because labels were deterministic functions of input features. Restructured the pipeline and shipped an honest model.
System architecture
Multi-user music coordination platform where the core problem is recommendation and personalization across heterogeneous content sources. FastAPI backend orchestrating specialized micro-agents for fairness reranking, taste synthesis, and cross-device sync across Spotify and YouTube. Abstracts two different music APIs into a single unified queue across 7 devices. Deployed on Railway and Vercel.
Tech stack
Available immediately
STEM OPT with 3 years of work authorization. No sponsorship required.
Boston, MA. Open to relocation and remote.