The AI-Native Product Manager.

Bridging the gap between product strategy and technical execution. Writing requirements, architecting system design and building the prototypes that prove them.

From multi-agent orchestration to RAG pipelines — this is my living lab of AI experiments, workflow automations, and shipped products.

50+ DAU 99.9% Uptime

Flagship Project

The Tenant's Voice: Engineering a Production-Grade RAG Pipeline

Scaling from prototype to 50+ daily active users with 99.9% uptime.

A dedicated, reliable platform that empowers UK tenants to navigate complex legal systems. Built to differentiate from general-purpose AI through verified sources and actionable guidance.

View Full Case Study

RAG Pipeline Architecture

User Query

+ Chat History

Query Transform

AI Keyword Extraction

Embedding

text-embedding-004

Vector Search

pgvector (Top 5)

LLM Generation

gemini-2.5-flash

Response

Structured JSON

Decision Log: Critical Optimizations

The initial system had 40s+ response times—unacceptable for real-time chat. I diagnosed the bottlenecks and made three key product decisions.

1. Model Selection

Gemini Pro offered +8.5% precision at 3x latency. Chose gemini-2.5-flash—3.8% hallucination rate acceptable with source citations.

2. Filter First, Search Second

Corrected index mismatch (L2 → Cosine) and added GIN keyword pre-filtering. 24x faster (911ms → 37ms).

3. Removed Redundant AI Step

Eliminated 27.7s pre-processing step. HITL testing confirmed vector-only search matched hybrid quality. 41s → 3.8s

RAG Optimization & Database Indexing

Problem: The match_documents function was responsible for 68.2% of query time. Mean: 911ms, Max: 2,620ms.

Root Cause: Index mismatch (L2 vs Cosine) and no pre-filtering across 6,000 documents.

Fix: Corrected vector index to vector_cosine_ops, added GIN index on keywords, implemented two-stage filtering.

Result: 24x faster (911ms → 37ms), max time dropped 66x.

Solving the 36k-Byte Crash

Problem: Most engaged users (long conversations) experienced app crashes with 400 Bad Request.

Root Cause: text-embedding-004 has 36,000-byte limit. Sending entire chat history was fatal for long chats.

Fix: Created dual-history approach: recentHistoryText for embeddings, fullHistoryText for LLM context.

Result: Eliminated crashes for power users, maintained conversational coherence.

Chatbot Optimization & Stability

Problem: 500 errors when AI returned plain text instead of JSON. End-to-end latency exceeded 40 seconds.

Root Cause: No try-catch on JSON.parse(). Redundant AI pre-processing step took 27.7s.

Fix: Added error handling with fallback message. Removed pre-processing—vector-only search proved equally accurate via HITL evaluation.

Result: 90% latency reduction (41s → 3.8s). Zero quality degradation.

Fixing Accidental Submissions

Problem: Users trying to write multi-line answers hit Enter and submitted partial thoughts, confusing the AI.

Root Cause: Single-line input box submitted on Enter—fighting user intent.

Fix: Replaced with auto-resizing textarea. Changed submit to Ctrl+Enter. Updated placeholder text.

Result: Aligned UI with natural workflow, reduced repeat questions.

AI Architecture & Agents

Experimental tools and agentic workflows built to solve real problems. Click any card to learn more.

The Green Room - Multi-Agent Sequential Debate System

Solving LLM "Echo Chambers" via Sequential Debate

Architected a sequential multi-agent system (Innovator → Risk Officer → Moderator) to force adversarial analysis.

Gemini 2.5 Google AI Studio

View Details

Autonomous Job Search Agent

Built an agent that reasons rather than just scrapes. Uses Semantic CV Matching to filter jobs.

90% Time Reduction 5x Daily Volume

Relay.app Google Sheets Gemini API

View Details

Image-to-Text-to-Image Pipeline

Built a pipeline to "translate" visual ads into detailed text prompts for precise creative iteration.

GenAI Pipelines Gemini

View Details

The Product Architect - Zero-to-MVP Chain-of-Thought Workflow

The "Zero-to-MVP" Chain-of-Thought Workflow

Automated PRD-to-Code using a 3-stage prompt chain with Router Logic for different coding tools.

2 weeks → 45 mins

LLMOps Prompt Chains

View Details

Before

After

Functional Feature

Rapid Prototyping Lab

From HTML to Functional Spec in <10 Minutes

Bridging the gap between Product and Engineering. I don't just write tickets; I take live HTML, inject AI-generated logic, and hand engineers a working functional prototype.

This eliminates the "Gulf of Specification" and validates technical feasibility instantly.

Vibe Coded Cursor Lovable

The AI-Native Product Manager.

The Tenant's Voice: Engineering a Production-Grade RAG Pipeline

RAG Pipeline Architecture

Decision Log: Critical Optimizations

RAG Optimization & Database Indexing

Solving the 36k-Byte Crash

Chatbot Optimization & Stability

Fixing Accidental Submissions

AI Architecture & Agents

Solving LLM "Echo Chambers" via Sequential Debate

Autonomous Job Search Agent

Image-to-Text-to-Image Pipeline

The "Zero-to-MVP" Chain-of-Thought Workflow

From HTML to Functional Spec in <10 Minutes

The Green Room

The Problem

The Solution

Key Insight

AI Job Application Co-Pilot

The Problem

The Solution

Results

Multimodal Content Engine

The Problem

The Solution

Key Differentiation

The Product Architect

The Problem

The Solution

Key Feature: Router Logic