Kareem Ghazal
Open to work · London / remote

Kareem Ghazal

I build software and AI systems, and deliver programmes. End to end.

20+ years across fintech, payments and financial services — VP at JP Morgan, CTO of a regulated fintech, PRINCE2 project / programme manager. I build and ship production software and applied-AI (LLM) systems end to end — with real eval harnesses, not vibes — and lead complex delivery. Most recently I shipped BTCBitByBit single-handedly across web, iOS and Android, and a growing set of live AI tools.

See the work Get in touch
GitHub LinkedIn Email
Selected Work

Things I’ve built

Live, working products — not just repos. A flagship fintech product shipped solo across three platforms, and a growing set of small AI tools.

Active Project

BTCBitByBit

Consumer fintech platform · self-custody wallet · web, iOS & Android

A consumer fintech product (Bitcoin education + self-custody) — and a real engineering build. At its core: a multi-protocol self-custody wallet (Lightning BOLT11 + BOLT12, Liquid, on-chain) on the Breez SDK, plus an on-device web wallet running the SDK in the browser via WebAssembly, so users hold their own keys straight from a tab.

Shipped end-to-end across web (Next.js / React / TypeScript), native iOS and native Android, on AWS / Terraform — designed, built, tested (Jest + Playwright) and operated solo, deployed to production daily.

Role: Architect, Lead Engineer & Product — solo, end to end

Python Django Next.js Swift / Kotlin AWS Terraform PostgreSQL Redis Lightning Network GitHub Actions CI/CD WAF ECS Fargate
Visit BTCBitByBit KayGee.Tech

AI tools I’ve shipped

A suite of small, live, open-source AI tools — each built around one idea: measure LLM systems, don’t vibe them, so every one ships a real eval harness. Grouped by area below — expand any category to explore. Built with FastAPI + Next.js on the Anthropic Claude API.

Patents & Legal Tech 5 tools A hallucinated fact here is malpractice — the model never decides the fact. ▾
ANTECEDENT BASIS

claim-check

A patent-claim structure & antecedent-basis linter — a deterministic engine decides the §112 drafting defects, and the LLM only explains each and suggests a fix. Recall / precision / grounding evals, plus a judge that keeps it to drafting help.

Engine decidesLLM only explainsNo legal overreach
Live demo ↗ Code ↗
PRIOR-ART CHART

claim-chart

Maps a patent claim’s limitations to supporting text in a prior-art reference (a claim chart) and flags what isn’t disclosed — a grounding filter drops any ‘disclosed’ quote that isn’t verbatim in the reference. Planted-disclosure evals.

Claim → prior artNo hallucinated disclosureNovelty verdict
Live demo ↗ Code ↗
CPC CLASSIFICATION

cpc-classifier

Classifies an invention into CPC patent codes via retrieval — the LLM only picks from retrieved candidates, so it can’t invent a code, and abstains when nothing fits. Top-1 / recall@k + no-hallucinated-code evals.

Retrieve then selectNo invented codesAbstains
Live demo ↗ Code ↗
PROSECUTION

office-action

Drafts a patent office-action response — a deterministic parser decides which limitations the rejection actually charted (verbatim-quote-verified), the LLM drafts the argument & amendment, and a no-new-matter gate verifies every added word traces to the specification. Includes the real Amazon 1-Click claim.

Parser decidesNo-new-matter gateReal granted claim
Live demo ↗ Code ↗
CONTRACT REVIEW

contract-lens

Reviews a contract or terms-of-service and flags the risky clauses (auto-renewal, liability, IP assignment, data rights) with a verbatim quote — grounded, with planted-clause evals. Educational, not legal advice.

Flags risky clausesQuotes the documentPlanted-clause evals
Live demo ↗ Code ↗
Clinical & Healthcare AI 5 tools Deterministic logic decides the clinical fact; the LLM only drafts or suggests. ▾
MONITORING & RECALL

care-recall

A safe demo of NHS primary-care monitoring & recall: a deterministic rules engine decides what’s overdue for a synthetic patient, and the LLM only drafts the recall — with recall / precision evals.

Rules engine decidesLLM only draftsClinical evals
Live demo ↗ Code ↗
CLINICAL RAG

monitoring-guide

Cited Q&A over drug/condition monitoring guidance that abstains when the corpus doesn’t cover the question — retrieval, citation-validity and faithfulness evals.

Clinical RAGAbstains, not guessesCited
Live demo ↗ Code ↗
SNOMED CODING

snomed-coder

Maps a clinical note to SNOMED CT codes via retrieval — the LLM only picks from retrieved candidates so it can’t invent a code, evidence must be quoted verbatim, negation & family-history mentions are filtered, and the SNOMED hierarchy subsumes parent codes. Abstains when nothing fits.

Retrieve then selectNegation-awareHierarchy-aware
Live demo ↗ Code ↗
RULE ROLLOUT

rule-shadow

Ship a clinical rule change safely: shadow-runs the proposed ruleset against the baseline over the same synthetic cohort, diffs the flagged volumes, gates on a golden regression cohort, and alarms when monitoring silently vanishes — you can’t A/B test on patients.

Shadow runGolden-cohort gateVolume-delta alarms
Live demo ↗ Code ↗
QOF REGISTERS

qof-register-builder

Builds NHS-style disease registers from SNOMED-coded records and flags what’s keeping patients off them — miscoded free text (with the exact suggested code) and status conflicts — deterministically; the LLM only drafts the completeness report, count-verified against the computed results.

Registers from codesMiscoding flagsCount-verified report
Live demo ↗ Code ↗
Agents, RAG & Data 3 tools Retrieval and tool-use, grounded in sources the system actually saw. ▾
TOOL-USE AGENT

research-agent

A tool-using agent that researches a company into a sourced interview brief — an orchestrated search + fetch loop, hard guardrails, and source-grounding evals.

Agent loopTool useGuardrails
Live demo ↗ Code ↗
CITED RAG

rag-doc-qa

Cited document Q&A that abstains when the docs don’t cover it — retrieval metrics, citation-validity checks, and an LLM faithfulness judge.

RAGEmbeddingsFaithfulness judge
Live demo ↗ Code ↗
TEXT-TO-SQL

text-to-sql

Turns a plain-English question into a read-only SQL query, runs it on a sample database, and returns the rows — measured by execution accuracy (did it return the right rows), not string similarity.

Runs the SQLRead-only guardExecution-accuracy evals
Live demo ↗ Code ↗
Code, Dev & System Design 4 tools Reviews real code and designs against grounded, measured criteria. ▾
CODE REVIEW

ai-code-reviewer

Reviews a diff or PR for real bugs and security issues — not style — with every finding pinned to a changed line and scored against a planted-bug test set.

Grounding filterPlanted-bug evalsPrecision
Live demo ↗ Code ↗
REPO ROAST

roast-my-repo

A witty-but-fair roast of any public GitHub repo — funny lines, a grade, and 3 genuinely actionable improvements, all grounded in the repo’s real README & facts, with a fairness judge.

Reads the real READMEFunny but fair3 real improvements
Live demo ↗ Code ↗
DESIGN REVIEW

arch-review

An AI architecture/design reviewer — paste a design or RFC and get severity-ranked findings across scalability, reliability, cost, security & operability, grounded in the design. Planted-weakness evals.

Reviews the designGrounded findingsRecall / precision evals
Live demo ↗ Code ↗
DECISION ADVISOR

adr-advisor

An architecture-decision trade-off advisor — weighs your candidate options across the trade-off axes and recommends one of them (never an invented one), with the trade-off it’s accepting. Decision evals.

Trade-off analysisRecommends only your optionsDecision evals
Live demo ↗ Code ↗
Pipelines, Evaluation & Safety 4 tools The plumbing of reliable LLM systems: structured pipelines, evaluation, and red-teaming. ▾
CV TAILORING

ai-job-copilot

A grounded LLM pipeline that tailors a CV to a job description — forced structured output, plus deterministic and LLM-as-judge evals.

Structured outputEvalsPDF export
Live demo ↗ Code ↗
PROMPT EVAL

prompt-lab

A tiny evaluation harness as a product — run a prompt over test cases, an independent judge scores pass/fail, and get a pass-rate, the failure pattern, and a better prompt. Its own eval proves it discriminates.

Prompt eval harnessIndependent judgeSuggests a better prompt
Live demo ↗ Code ↗
INTERVIEW PREP

interview-gauntlet

A mock interviewer that scores your answers against a rubric (relevance, depth, clarity, correctness) with feedback and a sharper follow-up — the grader is eval-tested to rank strong answers above weak.

Role-specific questionsRubric scoring + feedbackGrader is eval-tested
Live demo ↗ Code ↗
RED-TEAMING

llm-red-team

An autonomous red-teamer that adaptively attacks an LLM you own, detects breaches with an independent judge, and reports how to harden — recall / precision evals included.

Adaptive agentPrompt-injectionIndependent judge
Live demo ↗ Code ↗
Running & Fitness 3 tools AI for a coaching app runners trust — rules decide the plan & the analysis, the model only writes it up. ▾
TRAINING PLANS

run-coach

A running training-plan engine (a mini-Runna) — training-science rules decide a safe, progressive plan with a taper and Riegel paces, and the LLM only drafts the athlete briefing (easing off when you’re not 100%). Safety-gated evals + an opus faithfulness/safety judge.

Engine decides the planLLM only draftsSafety-gated evals
Live demo ↗ Code ↗
RUN ANALYSIS

run-insights

Paste a run’s splits → split, pace-fade and HR-decoupling analysis + coaching feedback that can only cite the real numbers — a grounding filter drops any hallucinated stat. Planted-pattern + grounding evals.

Split · fade · HR driftNo hallucinated statGrounding filter
Live demo ↗ Code ↗
TEXT-TO-SQL

run-query

Ask about your training history in plain English → it writes a read-only SQL query, runs it on a sample activity database, and returns the rows — measured by execution accuracy (did it return the right rows), not string similarity.

Runs the SQLRead-only guardExecution-accuracy evals
Live demo ↗ Code ↗

Profile

I'm a senior software engineer and PRINCE2 project / delivery manager with 20+ years across fintech, payments, capital markets and regulated financial services — JP Morgan Chase, Dun & Bradstreet, Virgin Media O2 and Tesco Mobile — delivering multi-million-pound outcomes, regulated products, complex programmes and teams I led. I move fluently between building software end to end and leading project, programme and change delivery — with governance, stakeholders and vendors. Most recently I built and shipped BTCBitByBit end to end on my own across web, iOS and Android — designed, built, tested (Jest + Playwright) and operated solo, shipped daily. I'd rather show you than tell you: the work is at btcbitbybit.com.

English & Arabic (Bilingual) London, UK
MSc Advanced Computer Science
University of Manchester
BSc Computer Engineering
UC San Diego
Cum Laude Honours · GPA 3.65/4.0
Prince2 Practitioner ISEB Diploma in Business Systems Development ISEB Software Testing Foundation

Impact in Numbers

0
Platforms shipped solo
web · iOS · Android
0
Live AI tools shipped
each with an eval harness
0
Annual savings delivered
JP Morgan Chase
0
First-year revenue generated
Tesco Mobile
0
Additional revenue captured
Tesco Mobile
0
Customers served
MyAhmed
0
Global team members led
MyAhmed
0
European markets migrated
Dun & Bradstreet

Experience

2023 — Present
KayGee.Tech Ltd
Architect, Lead Developer & Product Manager
  • Built btcbitbybit.com end-to-end across 3 channels: responsive web (Next.js), iOS app (Swift), and Android app (Kotlin)
  • Designed and built multi-account AWS infrastructure with multi-AZ HA, WAF, auto-scaling, and CI/CD via GitHub Actions OIDC
  • Secured UK copyrights for BTCBitByBit and Bitcoin Bit By Bit
2023
Dun & Bradstreet
Senior Programme Manager
  • Led AWS to GCP cloud migration programme across 10+ European markets
  • Defined future-state architecture vision and phased migration approach
2022 — 2023
Virgin Media O2
Senior Programme Manager
  • Led B2B data migration for the VMO2 merger across Salesforce, Eloqua, and Netcracker
  • Unified multi-system customer data onto a single platform, minimising business impact
2019 — 2022
MyAhmed Ltd
Co-Founder & CTO
  • Built and launched a mobile fintech platform serving 7,000+ customers
  • Led a 25+ person global remote team across 12 disciplines
  • Secured FCA Sandbox participation and AISP licence
2017 — 2019
Tesco Mobile
Programme Manager / Senior PM Digital
  • Launched Tesco Mobile for Small Business — £5M first-year revenue
  • Improved NPS from 5.9 to 7.2 and generated £1.1M additional revenue
JP Morgan Chase
Project Manager VP — Banking Regulatory Technology
  • Spearheaded global Sanctions Screening platform — £2M+ annual savings
  • Directed international expansion (Norway, Saudi Arabia)
  • Managed cross-functional stakeholders across Front Office, Risk, Payments, Trading, and Finance
Royal Bank of Scotland Group
Software Developer → Technology Project Manager
  • Progressed from graduate developer to project manager leading cross-platform projects
  • Rolled out global HR Payroll platform across the entire RBS group
Cimteq Ltd
Java Developer
  • Software development using Java

Skills & Expertise

Applied AI & LLM Engineering

LLM Applications RAG & Embeddings Agents & Tool Use Structured Output Evals & LLM-as-Judge Prompt Engineering AI Red-Teaming Anthropic / Claude API

Languages & Frameworks

Python Java Kotlin JavaScript TypeScript SQL React Node.js Django

Cloud & Infrastructure

AWS GCP Azure Docker Kubernetes Terraform PostgreSQL Redis

Engineering Practices

Agile / Scrum JIRA Git REST APIs Microservices CI/CD TDD / Testing

Project & Delivery Management

PRINCE2 Change Management Programme & Project Delivery Governance & PMO Stakeholder Management Risk & RAID MS Project Vendor Management

The Builder's Toolkit

Engineering is the core — but taking a product to market solo meant wearing every other hat too

Frontend Development (Next.js) iOS Development (Swift) Android Development (Kotlin) IP & Copyright Protection Brand Building Social Media Strategy Content Creation Video Editing (CapCut) Graphic Design (Canva) Community Management
Fintech Payments Capital Markets Financial Services Banking Regulatory Telecoms

Work With Me

Two decades of building, shipping and leading — most recently an entire product on my own, end to end, plus a growing set of live, open-source applied-AI tools. I'm open to senior software and applied-AI / LLM engineering roles, and to project / delivery management — contract or permanent, where I can be hands-on and help a team ship. Based in London, open to remote and to relocation.

🛠️

Hands-on Engineering

🧩

Architecture

🚀

End-to-End Delivery

👥

Technical Leadership

Let's Talk

Get in Touch

Interested in working together or have a project in mind?

I'd love to hear from you.

Kareem Ghazal

London, United Kingdom

kareem@kaygee.tech