Product · AI Systems · Built from Scratch

I don't just spec it.
I ship it.

Every project here started with a real problem, a team drowning in meeting transcripts, a compliance team missing regulatory signals, a chatbot that couldn't tell a founder from an analyst. I found the gap, designed the solution, and built the thing.

Claude API ◆ 0→1 Products ◆ RAG Systems ◆ TypeScript ◆ LLM Evaluation ◆ Python ◆ Regulatory AI ◆ Next.js ◆ Signal Intelligence ◆ Prisma ◆ Role-Aware AI ◆ Google Drive API ◆ Claude API ◆ 0→1 Products ◆ RAG Systems ◆ TypeScript ◆ LLM Evaluation ◆ Python ◆ Regulatory AI ◆ Next.js ◆ Signal Intelligence ◆ Prisma ◆ Role-Aware AI ◆ Google Drive API ◆

The Through-Line

Building AI products that work in the real world

Most AI demos collapse the moment they touch a real workflow, inconsistent outputs, no escalation logic, no way to audit what the model actually did. I build AI systems designed to survive contact with reality. That means thinking about failure modes before launch, not after.

I come to every product as both the PM and the person who can build a working prototype by Tuesday, which changes what questions you ask and how fast you move.

Across these projects, I've worked the full AI product surface: defining the use case, architecting the system (retrieval, context management, role-aware prompting), shipping the interface, and measuring reliability in regulated environments where silent failures are liabilities.

◆What I bring to an AI PM role

🧩

Product Definition

I find the real problem under the stated one. Every project here started from a workflow gap, not a feature request. I write specs, define success metrics, and kill bad ideas before they become bad code.

⚙️

Technical Depth

LLM context windows, retrieval pipelines, role-aware prompting, async message queues, I understand the stack well enough to call out engineering tradeoffs in a design review, not just in a post-mortem.

📐

Ship Velocity

All of this was built outside a big org. No platform team, no infra support. That forces ruthless prioritisation and an allergy to scope creep. I know what a v1 needs to be and what it can be later.

🏛️

Regulated Domains

Fintech, compliance, and enterprise AI are my home turf. I understand why a CISO blocks an LLM rollout and what it takes to make an AI product trustworthy enough to deploy inside a bank or regulator.

01All Projects

Filter by

TypeScript

2026

Kinage Intelligence

AI Product · Semantic Signal Infrastructure

Semantic intelligence pipeline for analyst-grade market signal classification, from raw feed noise to trust-weighted, ranked intelligence.

Outcome

Turned a daily manual process into a real-time, structured feed that analysts can filter and act on without leaving one interface.

Next.jsTypeScriptVercelWeb Scraping

AI Product · Behavioral Trust Infrastructure

Behavioral trust infrastructure for AI-assisted decision-making, role-aware constraint architecture with full audit trail for accountability.

Outcome

Role-aware AI responses that match user context, with async reliability and a full audit trail for accountability.

JavaScriptExpressClaude APIPrisma

AI Product · Compliance Decision Infrastructure

AI-assisted anomaly detection and risk classification for regulated institutions, surfacing what matters, suppressing what doesn't, with defensible audit trails.

Outcome

A purpose-built intelligence layer for compliance teams, designed to surface what matters and suppress what doesn't.

Next.jsTypeScriptVercelRegulatory Data APIs

Read breakdown →

Python

2026

Transcript Intelligence

AI Automation · Structured Knowledge Extraction Pipeline

Automated knowledge extraction pipeline, schema-constrained LLM orchestration that converts raw meeting transcripts into structured, actionable intelligence.

Outcome

Eliminated the manual meeting-notes process entirely. Structured intelligence is available in Drive within minutes of a meeting ending.

PythonClaude APIGoogle Drive APIGoogle Docs API

Read breakdown →

Python

2026

LLM System Reliability

AI Infrastructure · Evaluation & Drift Detection

LLM evaluation framework for regulated environments, drift detection, regression harnesses, and failure-mode classification calibrated to compliance-grade tolerances.

Outcome

A practical reliability layer that surfaces LLM degradation before it becomes a user-facing or compliance problem.

PythonLLM EvaluationMLOpsDrift Detection

Systems Engineering · Retrieval Architecture

Understanding retrieval latency at the systems level, before recommending it at the product level.

Outcome

First-principles understanding of the performance tradeoffs in retrieval systems that informs every RAG product decision I make.

C++Vector SearchIndexingRetrieval Systems

AI Backend · Decoupled Intelligence Orchestration

Decoupled AI orchestration layer, contract-first API design separating intelligence reasoning from product interface for independent deployment and iteration.

Outcome

A modular AI layer that lets the Kinage intelligence dashboard evolve its UI without touching the reasoning logic, and vice versa.

PythonAI / MLREST APISignal Processing

Ops Tooling · Incident Management Interface

Coordination overhead compounds outages. This removes it.

Outcome

Faster mean-time-to-resolution by cutting the coordination overhead that compounds every incident.

HTMLJavaScriptIncident Management