IniakiBoudiaf

Languages

Python, TypeScript, JavaScript, SQL, Java, R, MATLAB, HTML, CSS

AI & ML

LLM applications, RAG, MCP, prompt design, NLP, neural networks, fraud detection, model evaluation, AWS Bedrock, LlamaIndex

Frameworks

React, Next.js, Tailwind CSS, FastAPI, Hono, Node.js, Flask, Django, Supabase

Tools

Git, GitHub Actions (CI/CD), Playwright, Docker, Linux/macOS, Vercel, AWS, GCP, PostHog, Sentry, k6

Data

PostgreSQL, MongoDB, SQLite, Pandas, NumPy, scikit-learn, TensorFlow, ECharts, D3.js

Ways of working

Agile, Team Collaboration, Mentoring, Code Reviews

AI Personalisation Engineer

• Built AI systems used daily by the CEO and his team, including a private search and memory layer over company communications and a proactive agent that drafts work for approval.
• Improved retrieval accuracy from 86% to 92.6% by diagnosing a ranking flaw.
• Rebuilt an automated meeting coach as a service with a modelled 93% cut in LLM cost.
• Shipped product features and new apps to the Fanvue App Store, and load tested production.
• Built creative apps and a ten-game Minigames MVP for the new business models team.
• Automate QA workflows with Playwright and CI on GitHub Actions.
• Mentor colleagues through AI office hours and reusable AI tooling.

Data Engineer

• Engineered a crucial product demo for early-stage fundraising, instrumental in securing investment.
• Built a feature-rich B2C financial application using React, with AirWallex payments and real-time alerts.
• Developed key components for AIMA, an LLM-based RAG system, using Python, FastAPI and AWS Bedrock.
• Took on spikes to explore new possibilities, such as explaining parameters within the B2C app.

Junior ML Engineer

• Applied NLP to extract sentiment from financial tweets and generate signals.
• Built and tested text-to-SQL models and a fraud detection prototype on Ethereum data.
• Prototyped frontends for ML models in Flask and Django, accelerating user feedback cycles.

01

A private memory layer for an executive

PostgreSQL, pgvector, TypeScript, MCP, Claude

• Built a local-first, privacy-first retrieval system over email, chat, calendar, meeting transcripts and notes, exposed to AI assistants as cited, searchable MCP tools.
• Designed hybrid search combining four retrievers with Reciprocal Rank Fusion and a local cross-encoder reranker, with least-privilege access enforced at the database level.
• Gated releases on a fixed retrieval evaluation. When it stalled at 86% hit@10, profiling traced the gap to a global recency-decay term pushing correct older documents out of the results. Removing it lifted hit@10 to 92.6%.
• Treated delivery as the finish line: the system went live on the recipient's own machine and defects found in real use were fixed in the open.

02

A proactive AI agent with human approval

Claude Agent SDK, Slack, TypeScript

• Rebuilt an AI chief-of-staff as a proactive agent that surfaces one task at a time, takes plain-English approvals, and runs approved coding work in cloud sessions.
• Safety by design: the agent can open pull requests but can never merge them, and irreversible actions sit behind explicit approval.
• Earlier iteration moved from five specialised agents to one orchestrator plus skills, which was simpler to run and to secure.

03

Meeting coaching, from prototype to reusable skill

FastAPI, Python, Claude, Slack, Notion

• Turned a prototype automation into a multi-user coaching system giving each participant feedback after meetings.
• Migrated from a no-code workflow tool to a self-hosted FastAPI service, with a modelled 93% reduction in LLM cost per meeting.
• Rebuilt it with meeting-type classification, editable per-person principles and memory that avoids repeating advice, then packaged it as a shareable Claude skill.

04

Shipping apps for creators

Next.js, TypeScript, Hono, PostgreSQL, PostHog

• Worked as an embedded product engineer in a small team shipping to a shared production codebase: authentication, batch image generation and a credits payment flow.
• Found a quality-assurance gate that was silently failing open and passing every generated image.
• Built and launched a creator-to-agency marketplace app largely solo, using an unattended overnight agentic build workflow, and documented the method for colleagues.
• Load tested the production read path with k6 and showed that a larger database connection pool reduced throughput, a diagnosis that avoided an expensive wrong fix.

05

Rapid product bets

Next.js, monorepo, SDKs, Vercel

• Turned two AI generation tools into standalone apps on a shared monorepo, through QA, design feedback and a free plan.
• Prototyped a minigames platform in under four weeks: a game SDK and shell, ten games and a funnel site with SEO, built to test whether the idea is worth scaling.

06

An agentic engineering workflow

Claude Code, Playwright, GitHub Actions, Lighthouse, Vercel

• Orchestrate parallel teams of AI agents on large pieces of work, for example separate workers per feature plus a dedicated challenger agent reviewing an experiment loop, then merge their branches with review-driven fixes.
• Take work from implementation to a merged pull request in one verified flow: full local CI, Lighthouse, Playwright and screenshots, squashed commits and confirmed preview and production deploys.
• Separate real failures from infrastructure noise, such as stalled queues and resource limits, so unnecessary changes are never pushed.
• Start from scoping, not code: structured question rounds, then a plan document and phased issues, built from requirements consolidated across tickets, docs, chat and design files.

07

Making AI useful to colleagues

Claude skills, MCP, evaluation

• Built a generator that derives a person's writing guide from their own sent messages, validated in a blind two-alternative forced-choice test rather than by eye.
• Ran AI office hours and built skills and agents for colleagues in other teams, so AI adoption did not depend on one person.

Willingness to cut

Prefer one orchestrator over five agents and one skill over four. Do less, better.

Verify before shipping

Run the full checks, then confirm the deploy. One green run is not proof, and flaky tests are investigated, not ignored.

Plan from real context

Scope with the people involved and pull requirements from the source material before writing code.

MSc Data Science

University of Hertfordshire

Sept 2021 – Sept 2022 · Distinction

Thesis: Forecasting Ethereum gas fees using statistical, ML, and DL models.

BSc Financial Mathematics

University of Hertfordshire

Sept 2018 – Aug 2021 · First Class Honours

Project: Portfolio optimisation using single and multi-variable techniques.