Building AI Products, One Working Prototype at a Time.

Problem: Duplicate citizen reports clutter city maintenance queues.

Approach: Combined vision prompt scoring with geo-radius clustering to group incidents automatically.

What I'd improve: Add offline video frame batching for low-connectivity patrol vehicles.

Problem: PMs miss subtle weekly funnel drops in raw dashboards.

Approach: Scheduled n8n workflow formatting analytics JSON into structured LLM digest prompts.

What I'd improve: Add direct drill-down links to specific user cohorts in Slack messages.

Problem: Generic explanations don't land the same way across different audiences.

Approach: Prompt chain that first identifies audience context, then generates and refines an analogy against it.

What I'd improve: Add a feedback loop so users can rate analogy quality and steer regeneration.

Problem: Manually logging every job application update is tedious and error-prone.

Approach: n8n workflow parsing Gmail via GPT-4o, matching rows by a stable ID column instead of fragile header-text matching.

What I'd improve: Debugged issues included trailing spaces in column headers and ARRAYFORMULA re-triggering. Next step is handling more inconsistent header formats automatically.

Problem: Competitive research context gets lost between sessions.

Approach: Function-calling agent backed by Tavily for live search and Supabase for persistent memory, streaming responses via SSE to a React frontend.

What I'd improve: Add source citation display directly in the UI.

Problem: Interior design mockups are slow and expensive to generate.

Approach: Chained a vision-capable prompt with GPT-Image-1 to produce styled room renders from user photos.

What I'd improve: Move from prototype-scale image generation to a more controllable, fine-tuned rendering pipeline. This is also a genuine long-term venture I'm pursuing beyond the portfolio piece.

Problem: PM backlog scoring in messy spreadsheets is cumbersome to present to teams.

Approach: Clean, dynamic web interface recalculating prioritization indexes instantly upon input change.

What I'd improve: Add CSV export and Jira backlog integration.

Problem: Incident response is a multi-step, multi-role process that's hard to demonstrate as a single-agent demo.

Approach: Chained four purpose-specific agents (Sentry for detection, Detective for diagnosis, Medic for remediation, Scribe for the incident report) into one pipeline, deployed on Streamlit Community Cloud with a custom HUD-style UI.

What I'd improve: Add configurable incident scenarios and reduce per run API cost so it can be shared more widely. Currently kept unlisted to manage API credit consumption.

Problem: Personal context is scattered across WhatsApp, email, and calendar with no way to query it as one coherent history.

Approach: Chroma vector store for retrieval, entity-filtered exact matching for "who said what" questions, Claude synthesizes cited answers. React frontend, FastAPI backend, deployed on Vercel + Render.

What I'd improve: Hit a real memory ceiling on Render's free tier from sentence-transformers pulling in the full PyTorch stack; swapped to fastembed (ONNX-based) to fit in 512MB. Next: real multi-source data beyond the demo dataset.