Agentic AI
What It Is
"Healthcare Feature Review Board" is an Agentic AI-powered tool for Product Managers in Healthcare simulating a cross-functional product review. Just submit a healthcare feature idea, and four specialist agents—product, clinical workflow, patient experience, and risk/security—evaluate it for you in parallel, then a synthesis agent reconciles their findings into one recommendation.
Launch ProjectHuman-In-The-Loop Collaboration
I architected this multi-agent system using Claude Code as the engineering partner, working one phase at a time against a 16-phase build plan. Total build time from idea to final polish: 20 hours, solo. I maintained a strict manual audit at every checkpoint—reviewing agent outputs, catching malformed-JSON failures, verifying provider failback—before any commit reached main.
Tech Stack
Python/Flask, orchestrated with LangGraph. GitHub version control with automated CI/CD to Railway. The workflow graph renders live via Mermaid; results stream over Server-Sent Events (SSE). UI/UX is my proprietary design.
The Engine (Qwen3-235B-A22B)
Qwen3-235B-A22B-Instruct-2507, a non-thinking Mixture-of-Experts (MoE) model, served via the Hugging Face Inference Providers router with ":cheapest" routing across multiple providers for automatic failback. Four specialist agents run in parallel, each returning strict, schema-validated JSON; a synthesis agent reconciles their output. Malformed calls get one automatic retry; per-agent token caps bound cost and latency.
Why This Model
Its MoE architecture delivers large-model reasoning at a fraction of the inference cost, it's ungated with no licensing friction, and it's served by multiple providers—which made the ":cheapest" failback routing possible.
Generative AI
What It Is
"Movie Matchmaker" is an AI-powered cinematic recommendation engine designed for anyone who loves movies. Simply input your favorite movies, actors, or directors—and choose your desired vibe and streaming platform—and the system validates your input, analyzes cinematic themes, and dynamically generates tailored recommendations complete with match percentages, synopses, cast details, and streaming availability.
Launch ProjectHuman-In-The-Loop Collaboration
I architected this solo project through an AI-augmented prototyping process, employing Gemini 3 Flash to accelerate full-stack logic bridging back-end micro-infrastructure with front-end user experience. I maintained a strict manual audit layer throughout—correcting recursive context-loss issues, environment configuration drops, and prompt validation gates—to guarantee real-world edge-case stability.
Tech Stack
Python/Flask micro-framework. Version control managed via GitHub with an automated CI/CD pipeline deploying to a cloud-native Railway environment. All UI/UX features are my proprietary responsive glassmorphism design.
The Engine (Qwen 2.5 72B)
Utilizes Qwen 2.5 72B via the Hugging Face API Router for enterprise-grade reasoning. To control generation variance and mitigate hallucinations, I mapped a user-controlled sampling Temperature Parameter (0.1 - 1.0) tied to dynamic prompt strategies, dynamically shifting system directions from strict, deterministic baseline matches to high-creativity indie predictions.
Why This Model
Qwen 2.5 72B provides massive world-knowledge coverage across global film datasets, exceptionally strong adherence to complex system prompts, and native ability to reliably output clean HTML structures without backtick artifacts or formatting failures.