Work

Production AI systems

What I shipped at SJ Innovation LLC: the problem, what I owned, the constraints, and the outcome.

Simple Therapy

Sara: production healthcare agent

SJ Innovation LLC · 2024 – Present

Product · simpletherapy.com

500+ active users

Problem: A healthcare assistant needed specialised retrieval, persistent personalisation, and safe tool use against internal services, not a single prompt.
Role: End-to-end AI architecture: orchestration, tools, memory, backend, auth, and tracing.
Constraints: Authenticated access (Auth0 / JWT / JWKS), no exposure of clinical context, tool execution only through internal APIs, and production observability for quality, safety, and latency.
Outcome: Sara serves 500+ active users. Specialized agents handle intent, context retrieval, and response generation, with MCP tools, Mem0 memory, voice, and Braintrust tracing in production.

Sara: production architecture

User

Chat and voice

↓ Auth0 / JWT

Orchestrator · Sara

Routing, structured outputs, guardrails

↓ specialised agents

Intent

Analysis

Retrieval

Medical context

Generation

Response

Memory

Mem0

Persistent, personalised context across sessions

Tools

MCP servers

Authenticated calls to internal healthcare APIs

FastAPIPostgreSQLDocumentDBBraintrust

Braintrust traces quality, safety, and latency. Memory and tools sit beside the agent loop, not inside a single prompt.

What I owned

Multi-agent architecture and orchestration (OpenAI Assistants / Chat Completions, structured outputs, guardrails)
MCP servers connecting agents to internal healthcare APIs for real-time, authenticated tool use
Long-term memory with Mem0: semantic search plus contextual reasoning across sessions
Voice-enabled interaction and Braintrust tracing for quality, safety, and latency
FastAPI services, Auth0 / JWT / JWKS authentication, PostgreSQL and Amazon DocumentDB

Stack

Multi-agent systemsMCPMem0OpenAI APIsFastAPIAuth0PostgreSQLDocumentDBBraintrustElevenLabs

Client work. Product site linked; implementation is proprietary and is described without clinical data or internal APIs.

ImagineCreate AI

AI content studio for commerce

SJ Innovation LLC · 2024 – Present

Product · imaginecreate.ai

Generate → publish

Problem: E-commerce teams needed product imagery and video inside one workflow (generate, edit, caption, publish), not a disconnected model playground.
Role: Generation workflows, media pipelines, and the React product surface from generate through publish.
Constraints: Multiple model providers had to sit behind one product experience. Real media (upload, background removal, resize, video ingest) had to run end to end, then publish to Shopify and social platforms.
Outcome: Shipped AI generation for commerce: product background replacement, text-to-image, AI image editing, and text-to-video, with an in-browser editor, OpenAI captioning, and direct publish to Shopify, TikTok, LinkedIn, Facebook, and Twitter.

What I owned

Generation workflows wrapping Stability AI/SDXL, Midjourney, GCP image models, OpenAI, and Pictory behind a single product experience
Express APIs and media pipelines (AWS S3/Lambda, Sharp, PhotoRoom) for upload, background removal, resizing, and video ingest
React generate → edit → caption → publish flow, including an in-browser image editor and OpenAI-powered captioning

Stack

ExpressReactAWS S3LambdaSharpPhotoRoomStability AI / SDXLOpenAIShopify

Client work. Product site linked; no public repository.

ICR Capital

Convertible and call-spread advisory

SJ Innovation LLC · 2024 – Present

Pricing + syndicate

Problem: Advisors needed to price convertibles and call spreads, run multi-step bids, and rank a syndicate in one workflow, with derivatives logic that could not live only in a spreadsheet or the UI.
Role: Core backend: bidding, syndicate ranking, finite-difference pricers, and Excel deal warehouse, delivered as JWT-secured APIs.
Constraints: Inspectable quantitative logic (grids, Greeks, implied vol), a 33-sheet Excel warehouse, dual PostgreSQL routing, and shared JWT auth with the Node.js frontend.
Outcome: Advisors can price, bid, and analyse deals in one system: call-spread bidding, syndicate ranking, convertible and call-spread pricers, FRED Treasury curves, and SHA-256 idempotent Excel ETL with GitHub Actions for dev and live.

What I owned

Multi-step call-spread bidding and syndicate-ranking system, plus a 33-sheet Excel deal warehouse as JWT-secured REST APIs
Derivatives logic in Python (NumPy/SciPy): Black-Scholes and Crank-Nicolson grids, SK10 vol-spread, Greeks, implied vol/credit spread, make-whole tables, greedy syndicate construction
FRED Treasury curves, dual PostgreSQL routing, SHA-256 idempotent Excel ETL, GitHub Actions (dev/live), shared JWT auth with the React/Node app

Stack

PythonNumPy / SciPyPostgreSQLNode.jsReactJWTFRED APIGitHub Actions

Client work. No public product URL.

ICR AI

Multi-LLM enterprise chat

SJ Innovation LLC · 2024 – Present

Multi-provider gateway

Problem: An enterprise team needed one chat surface over several model providers, with memory, file RAG, usage accounting, and SSO, not a single-vendor playground.
Role: Platform: multi-provider gateway, Assistants API integration, cost tracking, and the streaming UI with Azure AD SSO.
Constraints: OpenAI, Anthropic Claude, and Google Gemini behind admin-configurable keys; streaming chat; token and cost reporting; role-based access.
Outcome: Shipped ICR AI with streaming chat, conversation memory, custom assistants (threads, file RAG, Code Interpreter, function calling), tiktoken-based usage reporting, and a ChatGPT-style React UI with Azure AD SSO and RBAC.

What I owned

Multi-provider LLM gateway (OpenAI, Anthropic Claude, Google Gemini) with streaming chat, conversation memory, and admin-configurable models and API keys
OpenAI Assistants API: custom assistants, threads, file RAG, Code Interpreter, and function calling against PostgreSQL and meeting-summary tools
Token and cost tracking (tiktoken plus per-model pricing) with daily and monthly usage reporting
ChatGPT-style React UI with Socket.IO streaming, markdown/code highlighting, assistant builder, Azure AD SSO, and RBAC

Stack

OpenAIAnthropic ClaudeGoogle GeminiAssistants APISocket.IOReactAzure ADtiktokenPostgreSQL

Client work. No public product URL.

Industry CV (PDF)