BUC Populi - AI Grading You Can Trust
Claude drafts rubric-anchored grades from a structured rubric, rules check attendance, and every grade waits for a professor's approval.
M.Sc. Computer Science·Westcliff University Irvine·MBA·Bethesda University of California·M.Sc. E-Business·SGH Warsaw School of Economics
Biography
I put AI into products people already use and measure it before anyone relies on it. At MintCollect I built Ask MintCollect, a Claude agent on Amazon Bedrock that works through 32 tools over the company's API, turns every change into a proposal the user confirms, and passed 210 of 210 answers in its eval. At Bethesda University, Claude drafts rubric-anchored grades and a professor approves every one before it counts. The throughline is the same everywhere: let the model do the first pass, prove it with evals instead of vibes, and keep a human on anything with real stakes.
Skills
Years Experience
10+
Projects
28
Agent Tools Built
32
Services
An engineer who builds with AI end to end - wiring the model into the product, measuring it, and putting the checks around it before it ships.
Decide what AI should and should not do: where a model adds judgment, where a human stays in the loop, and how the UX makes the AI trustworthy.
Give a model real tools over a real API with the user's own permissions, and turn anything that changes data into a proposal the user confirms.
Prompting, tool use, semantic search with embeddings and structured outputs, with the guardrails that keep them reliable.
Prove the feature works before users do: eval harnesses, code checks plus a calibrated judge, and picking the model by eval instead of by feel.
Selected Projects
Production features where AI does real work - with quality you can measure
Claude drafts rubric-anchored grades from a structured rubric, rules check attendance, and every grade waits for a professor's approval.
A Claude agent with 32 tools over the company's API and an eval that passed 210 of 210, plus semantic search with embeddings so a collector finds a card by meaning when the name fails.
Writing
Notes on engineering, AI, and building products that scale
Fine-tuning isn't the prestige move and prompting isn't the lazy one. The choice comes down to consistency, volume, and whether you can express what you want in words.
The 'best' model is rarely the right one. Picking between a frontier model and a fast, cheap one is a product call about latency, cost, and how much judgment the task actually needs.
A prompt isn't a magic incantation-it's a specification for behavior. Treat it like one: be explicit, constrain the output, give examples, and version it.
How it starts
A 15-minute call
After that
A reply within a day
Based in
Long Beach, CA · remote friendly