Sushant Joshi
I build production AI systems that users trust enough to run without me.
Most recent: a RAG and agent platform on AWS Bedrock that raised agreement with expert decisions from 62% to 81% and cut each review from 2-3 hours to about 5 minutes.
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 IT services firm (10,000+ employees, Fortune 500 clients). 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, raising agreement with expert decisions from 62% to 81% (validated on 100 historical cases) and cutting review from 2-3 hours per case to about 5 minutes.
System architecture
Swipe sideways to see the full diagram.
Independently built a HIPAA-aligned clinical risk prediction platform. XGBoost + LightGBM ensemble built on 100K 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
Swipe sideways to see the full diagram.
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.
Certifications
Professional-level certification scoped for engineers with 2-3 years of experience; passed as a recent graduate. Covers the full LLM lifecycle: prompt engineering, fine-tuning, GPU acceleration, inference optimization (ONNX, TensorRT), deployment at scale, and production monitoring.
Tech stack
Available immediately
F-1 OPT, STEM-eligible (authorized to work now).
Boston, MA. Open to relocation and remote.