The question I keep coming back to: when do people trust AI more than they should, and can our benchmarks tell?

• SHARE Lab and Critical ML Lab, grad researcher (Waterloo, Sep 2026 to now). A systematic review of how LLM trust benchmarks treat the people doing the trusting.
• SHARE Lab, undergrad RA with Sharon Ferguson (Waterloo, May 2025 to Apr 2026). Experiments on when people over-rely on AI advice. First-author paper under review at CHI 2027, 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.

A systematic review needs a search anyone can run again, and Google Scholar can't give you one. So I built the tools for it, plus one for the paper at the end. All three are open source:

• OpenProceedings: exact Boolean search over the titles and abstracts of NeurIPS, ICLR, ICML, AAAI, AIES, FAccT and IASEAI. No stemming, and workshop papers are left out unless you ask for them. A saved search replays to the same set of papers, says what changed if it doesn't, and reports its own PRISMA exclusion counts.
• scholarmend: fills in the real venue, year and track that Google Scholar's RIS exports cut off, so screening works from facts. pip install scholarmend
• refaudit: checks a .bib against Crossref, arXiv and OpenAlex before you submit. It's built for the hard case, a DOI that resolves to a different paper than the one the entry claims. pip install refaudit

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.

Parley: two coding agents, one shared findings doc, and neither context leaves its machine.

When two developers need to compare notes on code neither of them can share, Parley lets their agents do it. The agents swap typed claims with pointers into each other's repos, peer to peer, and both humans watch a cited record build up as it goes. It works across harnesses, so Claude Code can sit on one side and Codex on the other. Right now two people can pair and exchange a finding end to end over a real connection. Private for now, built with two collaborators.

OneBit Agent: an autonomous agent on a 1-bit model, running on a laptop CPU.

It's a prototype to see whether a ternary LLM can drive an agent you'd actually use, with no GPU and no cloud. The weights are about 90% smaller than FP16, it runs offline, and after the download each run costs nothing. It runs a bounded think, act, observe loop over a small sandboxed tool set. Tool calls are grammar-constrained, which keeps a weak model on the rails. There are also multi-agent pipelines (planner, researcher, writer) that can switch to a stronger local model through Ollama when the job needs it. Private for now.

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:

• Responsible AI and VLA systems, for my master's
• Natural language to SQL, at Natural Schema
• Demand forecasting, at D&I Integrators

Before that it was RAG pipelines, squeezing models onto local hardware with llama.cpp, and conversational search.

Now

• University of Waterloo, MASc student and TA for MSE 121 (Sep 2026 to now). Thesis on responsible AI and VLA systems.
• 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, 18.7M units a quarter across 183 SKUs.

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 (WAT.ai), 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 with distinction 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, with a thesis on responsible AI and VLA systems. Co-supervised by Sharon Ferguson (SHARE Lab) and Sirisha Rambhatla (Critical ML Lab), with MSE 121 to TA on the side. Expected finish: August 2028.

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