CV

Seven years shipping production software. The last four building the AI and data layer of Shurq from idea stage to 200+ concurrent users, alongside client delivery in the US, Canada, the UK and the Gulf. I want to go deep on one product with one team.

Download the PDF Two pages. Same facts, laid out for print.

Selected delivery

Most of this is client work under NDA, so each entry describes architecture and outcomes rather than showing source. Where there is a full write-up, it starts with what went wrong.

  1. Shurq: data pipeline, MCP server and AI Ad Manager

    2022 to present

    UK Amazon-analytics SaaS, joined at idea stage, leading AI and data engineering

    Shurq is now a paid SaaS (v2.77, 660+ logged changes) serving Amazon brands across 21 marketplaces, with a 10+ person team and 200+ concurrent users. I own the pipeline that pulls sales, fee and ad data from Amazon every hour and turns it into true per-product profit; every dashboard, alert and agent reads from it. I led and built the MCP server: 1,080 tools, 29 of them write actions, over streamable HTTP with scoped API keys, usable from Claude, ChatGPT, Cursor or Claude Code. And the AI Ad Manager: rules, scheduled workflows and an approval queue, every change logged and reversible, with LLM fallback from self-hosted Ollama to OpenAI.

  2. Bilingual AI learning platform and school-network dashboards

    2025 to present, in pilot

    Québec EdTech client, raw requirements to two shipped products

    A learner platform that generates lessons, quizzes and reflections by grade, language and pace, with a Premium tier on Stripe, and a role-based dashboard suite for school service centres, schools, teachers and parents. Both EN/FR, data hosted in Canada. Reworking the lesson-image prompts and text handling cut text-rendering errors by roughly 80% across 8 EN/FR scenarios, with no model change and no added cost or latency. An EN/FR-parity Playwright suite blocks a release when either language regresses.

  3. Voice agents that act in the CRM mid-call

    2026

    US payment-services client (sole engineer), Canadian SEO SaaS

    Outbound calling platform with Pipedrive dedupe, deal creation, callback scheduling, 10-scenario follow-up email routing and an SMS tool that sends a document during the call. Compliance built in: dialling windows, DNC scrubbing, mid-call opt-out detection, verbal-consent logging, CASL unsubscribes. The early hang-ups traced to speech-to-text hearing “hello” as “no”.

    Read the case study

  4. Agents with write access to business systems, made safe

    2025 to 2026

    Croncore delivery agent, and leadflow-agent (public)

    The Croncore agent triages pull requests, tracks tasks and drafts client mail; every side effect waits in a pending_actions row for a human, and each run is scoped to one project. leadflow-agent turns an inbound enquiry into HubSpot records: a human approves each write, every action goes to a ledger with undo, and it costs under $0.002 per lead.

    Read the case study

  5. Arabic (Saudi-dialect) inbound sales agent

    2026

    Live on web and phone

    Answers in about 0.5 s, captures the lead and hands off to a human. Model fallback plus a scripted fallback means it never goes silent, and Twilio webhooks are signature-verified.

  6. Speech service into a legacy PBX

    2024

    Mongolian client, language with no commercial vendor

    Built and operated the serving stack: a self-hosted FastAPI service running two off-the-shelf engines behind one interface with fallback between them, producing audio that matched the client's existing Asterisk/FreePBX contract so nothing on the telephony side had to change.

    Read the case study

  7. Also built

    2025 to 2026

    Products and internal tools

    Offline-first retail POS (Tauri/Rust and React, ~380 automated tests). Live speech-to-speech translation with latency shown on screen. An MCP coordination server for coding agents. A prospect-research pipeline where nothing is sent without review. A multi-user time tracker on Cloudflare Workers and D1. Led a retrieval agent that quotes live prices from an automotive client's own data.

Experience

Integrated with

Each one wired up and debugged in production, not read about.

Amazon Ads & Selling Partner APIs · HubSpot · Pipedrive · Twilio · Asterisk/FreePBX · Deepgram · ElevenLabs · ClickUp · GitHub · Gmail · Google Sheets · Google Places · Stripe

Skills

Languages
Python · TypeScript/JavaScript · Node.js · SQL
Agentic & LLM
tool-using agents · MCP tool servers · Anthropic tool use · OpenAI function calling · schema-validated structured outputs · RAG · provider fallback · human-in-the-loop approval gates · prompt-injection containment · eval design and calibration
Backend & data
FastAPI · Django/DRF · Next.js · PostgreSQL · pgvector · Redis · Celery · BigQuery · hourly ETL pipelines · Docker · rate limiting and backoff
Cloud & ops
AWS · GCP · Azure · Vercel · Render · Cloudflare · GitHub Actions CI/CD · Prometheus/Grafana

Education and availability

BSc Software Engineering, PAF-IAST.

Remote, working across EU and US hours. Open to relocation where the role offers visa sponsorship (Ireland, the EU, the UK). Open as an employee or on contract.

Hiring an AI engineer for agents, integrations or data pipelines? Send me a real problem you are dealing with and I will tell you how I would approach it. That is faster than a screening call for both of us.

hello@sparli.dev