An AI agent for the Commercial team at a biopharmaceutical company.
Reps and analysts ask a question in plain English and get one answer drawn from documents, emails, meeting recordings, SQL data and Power BI reports.
The model never writes SQL. It writes a JSON request that my code validates and compiles, so model-written SQL can't reach the data.
95%+
retrieval accuracy, up from 68%
- Every page and slide read by a vision model
- Chunks split at headings, not word counts
- A context tag on every chunk
Seconds
to answer a lookup that used to take days
500+
documents, emails and recordings searchable in one place
Hiring for the world's largest school.
It grades every exam, practical and interview, and replaced a paper process that took 3 to 5 weeks. I built it alone, as co-founder and sole engineer, in five months.
Written answers go to a council of models that check each other. A person only sees the rare answer they can't agree on.
5,000+
candidates evaluated
- Results in under an hour after the exam ends
- Every interview video scored and checked for cheating
- Up to 1,300 candidates testing at the same time
3,500+
teacher-hours saved each hiring round
98%
video integrity accuracy, under 0.5% false positives
A multi-agent hiring platform built on verified work, not resumes.
Seven AI agents evaluate each candidate, from discovery to the final report. I built it alone, end to end.
LLM scores bunch together, so I designed Dynamic Reference Bucketing: it scores candidates through a few pairwise comparisons instead of absolute numbers.
99%
of the true score spread recovered, up from 42%
- About 2 pairwise comparisons per candidate
- 8× lower error than raw LLM scores
- 1.4% of pairs ranked wrong, down from 19.6%
5 min
per evaluation, down from a week
95.7%
agreement with recruiters across 23 pilot evaluations
About
- Education
Drexel University
BS Computer Science, 2023 to 2028
- Languages
- AI
- Web
- Data and cloud