AI Engineer

What I worked on

• •Shipped production agents in TypeScript on the company's agent SDK and internally hosted inference, including a sales-call analyzer, an interview analyzer, and the surrounding workflows such as an applicant tracking system.
• •Designed the agents around a multimodal pipeline rather than transcript-only prompting, fusing speaker diarization, acoustic features, and speech emotion recognition with semantic analysis so the system scores how something was said, not only what was said.
• •Built the sales agent to read intent past the literal answer, separating prospects who mean yes from those who agree on the call and cancel later, and to give reps call-over-call feedback on delivery.
• •Built the interview agent to score authenticity, clarity of thought, and engagement from vocal delivery, flagging candidates reading answers off a screen and cross-checking claims against resume experience.
• •Built the evaluation harness the agents are scored against: a golden set of human-labeled clips where continuous model outputs are bucketed into bands and a prediction counts as correct only when it lands in the human label's band, kept agent-agnostic through an adapter interface.
• •Engineered the harness so the keep-revert decision logic is pure and unit-testable without a live run, and so a flaky batch can never become or silently poison a baseline, with rejected runs still recorded so a week of pipeline collapse cannot read like a week nobody ran the eval.
• •Built an auto research loop that runs unattended for hours, mutating agent parameters, scoring each variant against the golden set, and keeping only what measurably improves, with a bounded-parallel runner replacing sequential execution.
• •Found the evaluation metric itself was gameable: a constant prediction of zero scored higher than the tuned model on every dimension. Halted tuning and invalidated the conclusions that depended on it rather than reporting the favorable number.
• •Root-caused chance-level accuracy to speaker diarization collapse rather than model quality, after an over-loose clustering threshold tuned on three clips merged all speakers on 8 of 13 clips and forced downstream scores to exactly zero. Retuning reduced collapse to 0 of 13 and raised band accuracy on the affected dimension from 0.15 to 0.46.
• •Automated regression gating in vitest so every model and version change is scored against held-out sets before release, replacing manual review and catching regressions before production.
• •Designed the UI and built payment endpoints, agent CRUD, and developer-facing logging and success-rate views for the marketplace, digital vault, and developer portal.