project file / jun–sep 2026
Arvya
Evidence-backed agent workflows for publishing, document generation, advisor research, and product delivery.
- my role
- AI engineering & GTM intern, Arvya
- status
- Product and pilot work
- field
- Applied AI / product engineering
- tools
- TypeScript · Python · Agent workflows · XLSX / PPTX / PDF / DOCX · SEC filings
01 / the problem
Why this exists
Financial workflows can produce polished output before anyone has checked the sources or the boundaries. At Arvya I worked on systems that keep evidence, review, and deliverables connected.
02 / the build
What I worked on
Built a social publishing agent that turns approved, public-safe company knowledge into LinkedIn and X drafts, routes editorial and visual review, and queues approved posts through Buffer.
Engineered financial-plan and investment-proposal workflows that create cited spreadsheet, slide, PDF, and document artifacts from a shared representation, with audits and durable jobs.
Developed a source-backed wealth-signal pipeline using public filings, and shipped advisor-facing dashboards and client deliverables. I also worked on production agent routes, analytics, and published product articles.
03 / engineering notes
Choices that mattered
- Generated claims carry source context and a public-safety decision before they reach a draft. Publishing remains a separate, explicit step.
- Document generation shares one intermediate representation so the spreadsheet, deck, and written output can be checked against the same facts.
- Research signals retain their source trail and pass through review before anyone acts on them.
04 / reflection
What I took away
The interesting engineering work lived at the handoffs: from evidence to claim, claim to artifact, and artifact to a person who could approve it. I describe the architecture here without exposing client records or internal data.