Garden Grove, California · open to Summer 2027
Full-stack engineer · AI systems
I build AI systems that have to hold up in front of real students — the agent, what it remembers, what it is not allowed to do, and the tests that prove it.
Signalslive
Real git history across the repositories I work in, under my four identities — not an embedded badge.
Aboutprofile
I'm the founding engineer at Someta, an AI tutoring platform used by about 300 students across three partner schools. I'm the only full-time engineer, so I own all of it: a FastAPI service of roughly 1,400 Python modules, the agent orchestration layer, the React and TypeScript client on top of it, and the roughly 6,300 tests that keep the whole thing honest.
Most of my work is on the AI systems. A tutor that guides a student to the answer and never hands it over. Compaction that extracts what is worth keeping before it compresses the rest. A safety check that lives outside the model it constrains, so a compromised prompt cannot suppress it.
The other half is the unglamorous part that makes the first half usable: a cost fix shipped alongside the ledger that proves it worked and the environment variable that reverts it, an auth flow that refuses a token it cannot verify instead of guessing, and the interfaces a teacher actually has to open between classes.
I came to this through community college — five associate degrees from Cypress while working full-time — and I'm now at UC Irvine, graduating June 2028.
Selected workcase_studies · 4
Agent architecture · 9 cognitive tools
Three-phase streaming, a tool library the orchestrator selects from rather than listing, per-agent scoping, and the continuation call that fired on every lesson opener and produced zero characters.
Compaction · memory · vector retrieval
Auto-compaction that extracts memories before it summarises, quality scoring on a cheap model, and Firestore native vector search as a single source of truth instead of a second database that can disagree.
AI safety architecture · K-12
A classifier upstream of the model it constrains, so a compromised prompt cannot suppress a check that already ran — plus a data boundary where structure crosses and raw text does not, and an eval harness so it is measured rather than asserted.
Inference cost · ~$20 → ~$10 per student
A 14–36K-token prefix re-sent at full price every turn, because a five-minute cache window is shorter than a student thinking. Fixed, reverted by one env var, and measured with a ledger I built to check my own claim.
Experienceexperience · 3
Sole full-time engineer across two codebases — 4,171 of 4,792 commits on the core platform and every commit on the Sage guide engine. Designed the LLM tutoring agent and its memory, retrieval and safety layers. Built Adaptive Assignments, the largest system in the repo, alone: a multi-tenant product for charter school districts with LEA-scoped tenant isolation and a three-tier PII model.
On the client side I built Stride, the teacher dashboard for grading, assignments, help requests and interventions, the unified student chat surface, and the design system behind both — roughly 255,000 lines of TypeScript and TSX I also maintain. Shipped the Canvas integration, and presented the technical work as the company raised $100K to launch v1.
Also written up: the Canvas LTI 1.3 launch — RS256 pinned, nonce and deployment bound, failing closed on an unprovisioned registration — and zero-downtime secret rotation with dual-key accept-both and a least-privilege Lambda.
On a four-person team for EY's sustainability practice. Ran a ~40-person survey analysed in SQL with two-sample t-tests, and presented four proposed solutions to EY mentors — the leading concept reached 80.5% opt-in.
Taught and graded mathematics and English for 60+ elementary students weekly.
Projectsprojects · 2
Lens watches a Claude Code session as it runs and explains each turn on-device, so you can follow what an agent is doing without shipping your codebase anywhere. A hotkey hands any screen region to a local Qwen model. Zero telemetry — inference never leaves the machine.
A visual, terse output mode for Claude Code: lexical compression plus ASCII diagrams, tables, trees and sparklines, so answers are faster to scan and verify at minimal token overhead.
Educationeducation · 2
B.S. Computer Science · Donald Bren School of ICS
Campuswide Honors Collegium · Chancellor's Excellence Scholarship
Five associate degrees, all conferred with High Honors
A.S.-T. Computer Science · A.S.-T. Mathematics · three A.A. Liberal Arts. Honors Program, President's Honor List ×3, and the college's first Data Club — all while working full-time.