
hi, Waleed here.
I build multi-agent systems, RAG pipelines, and LLM workflow automation.
My focus is the work most demos leave out: evaluation, observability, and the engineering that turns a working prototype into something a team can rely on every day.
Graduated from the University of Bologna in October 2026 with a Master's in Automation Engineering and a thesis built at MemorAIz S.r.l. Now open to AI and ML engineering roles.
About
Half full-stack engineer, half AI/ML. How I got here, and what I'm looking for next.
- 985Automated tests across the thesis service, its console and its first tenantRemindrAI
- 97.5%Item-match accuracy across six supermarket chains, against a 95% targetFruugle
- 99.5%Product-category accuracy across 16 classesFruugle
- 100K+Records moved through ETL pipelinesJubilee Life Insurance
Approach
I build models the way I would build any other production service: real APIs, error handling, observability. The alternative is a notebook that runs beautifully on my laptop and nowhere else, and I have written enough of those to know better.
A year rotating through cybersecurity, web development, and database administration means I can hand over a whole working system, rather than the ML half and a hopeful note about deployment.
Most of the work goes on the unglamorous part: retries, idempotency, and whatever decides to fail at three in the morning. Nobody puts that in the demo. It is usually the reason the demo still works next week.
Education
Master's in Automation Engineering
Alma Mater Studiorum - Università di Bologna, Italy
Oct 2024 - Oct 2026
Erasmus+ Mobility
University of Twente, Enschede, Netherlands
Sep 2025 - Jan 2026
Bachelor's in Electrical Engineering
NED University of Engineering and Technology, Karachi, Pakistan
Oct 2019 - Aug 2023
Roles I'm targeting
- AI Engineer
- Agents, tool use, and the infrastructure that makes them dependable.
- ML Engineer
- Models trained, evaluated, and served like any other production service.
- Full-Stack AI Engineer
- End to end: the model, the API, and the interface it is served through.
Experience
The roles I've held and the systems I built in each one.
AI Engineer · Master's Thesis
MemorAIz S.r.l.Feb 2026 - Sep 2026Bologna, Italy- Built RemindrAI for my thesis: a multi-tenant reminder backend, its operator console, and the Mastra capability other AI assistants register in one line.
- Made delivery retry-safe with an atomic database claim, per-channel idempotency, and database-owned backoff; median claim latency stayed near-flat from 1K to 100K due reminders.
- Engineered the control plane: per-organization API keys, tenant isolation at four independent depths, plans and quotas as data, and audited GDPR export and erasure.
- Shipped EU-pinned on Vercel under 985 tests across three repositories, with H-FARM's Student Assistant as the first production tenant.
Data Clustering & AI Model Intern
Fruugle SIAJan 2026 - Apr 2026Jelgava, Latvia (Remote)- Built an XGBoost model that matches the same grocery item across 6 Italian supermarket chains, reaching 97.5% accuracy against a 95% target.
- Trained a second model to classify products into 16 categories at 99.5% accuracy, using class weighting so under-represented classes were not ignored.
- Engineered the pipeline over 80K+ scraped products: Italian-text cleaning, a quantity parser, and hard-negative sampling for ambiguous pairs.
- Wrote an inference script that grouped 24K+ unsorted products, with guardrails to reject implausible matches.
Tech Trainee
Jubilee Life Insurance Co. Ltd.Jan 2024 - Oct 2024Karachi, Pakistan- Rotated through web development, databases, networking, and cybersecurity, building full-stack applications with React, ASP.NET Core, PHP, and MySQL.
- Built ETL pipelines over 100K+ records and tuned queries and stored procedures, cutting database response times by ~40%.
- Automated SharePoint and CRM workflows for internal teams, reducing manual effort by ~60% and speeding up client response.
- Monitored threats on IBM QRadar (SIEM/SOAR/PAM) and delivered 15+ SSRS reports and Power BI dashboards for business analytics.
Selected projects
Thesis and independent work across applied AI: agents, retrieval, evaluation and fine-tuning. Source code is linked wherever the repository is public, and a live demo wherever one runs.

Master's Thesis · MemorAIz S.r.l.
RemindrAI: Embeddable Multi-Tenant Reminder Capability
A reminder and messaging service that AI assistants embed instead of rebuilding. An agent turns plain chat into reminders delivered over Email, WhatsApp, Telegram, and Discord, and each reminder's owner is pinned outside the model, so a prompt injection cannot retarget it.
Highlight: Its first tenant adopted it with a base URL and a token, holding no scheduler, queue, or provider secret.
Private repository · MemorAIz

Independent Project
AI Workflow Assistant: Data Analyst for Spreadsheets
An AI data analyst for spreadsheets: upload a CSV or Excel file, ask a question in plain language, and get a report with metrics, charts, findings and recommendations. A LangGraph agent plans the analysis and calls deterministic pandas tools, and nothing the model writes is ever executed as code.
Highlight: Every number in a report is checked against the pandas computation behind it before you see it, backed by 508 tests.

Independent Project
Multi-Agent Research Team: Open-Source AI Research Platform
A research platform where three CrewAI agents search the web, write runnable code examples and deliver a cited Markdown report, while the browser shows every search, retry and provider fallback as it happens. Every report link is checked against the URLs the search actually retrieved, and the rest are flagged.
Highlight: Jobs survive restarts and deploys, and per-stage provider fallbacks keep the free-tier demo running, with 259 tests across pytest, Vitest and Playwright.

Independent Project
Self-Correcting Multi-Agent RAG
A RAG system that checks its own work: a guardrail agent drops off-topic chunks from hybrid BM25 and vector search, and an evaluator grades each answer against its sources, correcting weak ones in a bounded loop. A tool-using agent adds read-only SQL, and visitors can search their own uploads.
Highlight: Hybrid search lifted recall@5 from 0.967 to 0.992 over dense retrieval on a 60-question golden set, backed by 347 tests.

Independent Project
LLM-as-Judge: Evaluation Platform for LLM Outputs
A platform that grades LLM outputs with an LLM judge: pairwise, pointwise, reference-based, ranking, RAG faithfulness and agent trajectories. Each pair can be judged in both orders to offset position bias, and results flag suspected prompt injection, verbosity bias and self-preference.
Highlight: Published on PyPI as one engine behind a library, CLI, REST API and Streamlit dashboard, with 98% test coverage.

Independent Project
FastAPI Niche Assistant: Fine-Tuned Small Language Model
A 1.5B-parameter Qwen2.5-Coder model fine-tuned with LoRA to answer FastAPI questions, quantized to a 1 GB GGUF and served by llama.cpp on CPU behind an OpenAI-compatible streaming API. A Cloudflare Worker rate-limits requests at the edge in front of Google Cloud Run, all on free tiers.
Highlight: An eval gate decides each release: the Qwen fine-tune scored 0.666 against 0.546 for the TinyLlama release it replaced, on a 108-item held-out set.
GitHub activity
View profile (opens in a new tab)Skills
The tools and frameworks I work with, grouped by where they apply.
- AI / LLM Engineering
- ML & Data
- Languages
- Web & Backend
- Data & Messaging
- Cloud & DevOps
Contact
Whether it's a role, a question, or a collaboration, send a message. I read every one.