Natural Schema: ask your data in plain English.
You type a question. An agent writes a read-only query, puts it through a safety gate, runs it sandboxed, and hands back a table. The SQL comes with it, so you can check its work.
"What were our top 5 products by revenue last quarter?"
SELECT product, sum(revenue) AS rev
FROM sales
WHERE quarter = 'Q2'
GROUP BY product
ORDER BY rev DESC
LIMIT 5;
Nothing to write, nobody to wait on. Early access is opening in waves, so join the waitlist if you want in.
Retrieval, agents, and forecasting, mostly. I've shipped that kind of thing for industrial operators at Cognite, for clinicians at XCare, and into a social robot in Norway.
Three things have my attention right now:
- Natural language to SQL, at Natural Schema
- Demand forecasting, at D&I Integrators
- Responsible AI and VLA systems, for my master's
Before that it was RAG pipelines, squeezing models onto local hardware with llama.cpp, conversational search, and a couple of years of research on when people trust AI advice more than they should.
Now
- Natural Schema, co-founder (San Jose, Apr 2026 to now)
- D&I Integrators, ML engineer (Toronto, Jun 2026 to now). Demand forecasting for production planning, 20M units a quarter across 173 SKUs.
- University of Waterloo, MASc student and TA for MSE 121 (Sep 2026 to now). Responsible AI and VLA systems.
Research
- SHARE Lab, undergrad RA with Sharon Ferguson (Waterloo, May 2025 to Apr 2026). Experiments on when people over-rely on AI advice, first author on a CHI and CUI submission, and the study pipeline is public.
- IR group, RA with Mark Smucker (Waterloo, Jun 2025 to Apr 2026, then Jun to Aug 2026). Got popularity bias out of recommender evaluation, and reviewed where conversational search has got to.
Before
- PathAI, software engineer (Boston, Sep 2025 to Dec 2025). Built their first customer-facing cost dashboard, shipped in AiSight Dx v2.18.
- NTNU, social robotics research engineer (Trondheim, Jan 2025 to Apr 2025). Put a RAG agent inside a Furhat social robot, and the backend is open source.
- Cognite, software engineer (Austin, May 2024 to Sep 2024). Built the first agent to ship on their "Atlas" Industrial AI platform.
- XCare, founding AI engineer (Toronto, Oct 2023 to Oct 2024). X-ray classification with CNNs and ViTs, plus a RAG system for rehab guidance.
- Genellipse, software engineer (Toronto, Sep 2023 to Dec 2023). MongoDB architecture for vector similarity search across 13 collections.
- Approva Financial, full-stack developer (Toronto, Jan 2023 to Dec 2023). MERN app for a fintech going through Techstars.
- TD, automation QA engineer (Toronto, Jan 2023 to Apr 2023). Selenium test framework that ran 58% faster than what it replaced.
- Microsoft and UWaterloo WEA, project team (Waterloo, May 2022 to Aug 2022). Azure ML tool for sorting physical disability services.
The longer version is on LinkedIn.
AI / ML: PyTorch, HuggingFace, RAG, agents, VLA systems, OpenAI, Gemini, llama.cpp Backend: FastAPI, Django, Node.js, Docker, AWS/GCP/Azure, microservices Full-stack: React, Vue, TypeScript, Express, SQL/NoSQL Data: Azure ML, ChromaDB, Supabase, Jenkins, Firebase
- Search engine: BM25 plus embeddings, with query-biased summaries and a statistical evaluation
- Audio transcriber: Whisper and GPT cleanup, with a human in the loop
- Meal Stream: meal planning and nutrition analytics
- Energy price forecasting: ML models for the PJM market
- NBA player projection: a deployed model with a dashboard on top
BASc from Waterloo in Management Engineering, computing option. Machine learning, optimization, HCI, software engineering, systems design, decision analysis.
Now a MASc student there in Management Science and Engineering, on Responsible AI and VLA systems, with MSE 121 to TA on the side.
Brazilian Jiu-Jitsu, marathons, football (both kinds), basketball, squash, tennis, golf. I also collect colognes and watch absurdly long sports documentaries.
- Natural Schema, if you want to ask your data something
- LinkedIn and GitHub
- j29parma@uwaterloo.ca