Founding Software Engineer at Mando
I joined as a founding engineer and work across the React frontend, the Python services and the cloud.
Workflow capture. I built active and passive workflow capture. It pairs screen recordings with DOM events, so a capture keeps the application actions as well as the video. A first guide appears in seconds while full generation runs in the background, and the same capture feeds search answers with citations and no-code testing.
Video editor. I rebuilt the React video editor around one shared timeline for trimming, annotations, redaction and voiceovers. QA found that two tabs editing the same guide silently overwrote each other, so I added optimistic concurrency control and three-way merging. A partner’s 4K video with roughly 8,000 annotations hung the editor; indexing the annotations made it responsive again, and automated performance benchmarks now cover load, interaction, playback and save.
Pipeline reliability. I hardened the AI generation and editing pipelines: explicit state transitions, retries, output validation and fallback models. I fixed races in how edits propagated, replaced polling with WebSocket job updates, and grouped terminal failures by root cause so the team could see why a workflow failed or stalled.
Meeting bot and search. I built a cross-platform meeting bot end to end. It turns a recording into a summary, action items and a record you can search and share. I also extended search with streaming answers, source citations, screenshot attachments and links to the exact workflow step, both in the app and in the browser side panel.
Partners and managed services. I shipped branded apps and Chrome extensions for 3 partners. Partner deployments now run through one Terraform and GitHub Actions setup. I also built managed-services ticketing: requests route to the right specialist, SLA timers count business hours, and expert invoices go out automatically through Stripe Connect.
Data integrity. I wrote validation and repair tooling for stored workflows. It checks for invalid timestamps, missing sources and gaps in step order. Running it on production showed legacy asset links still pointing at DigitalOcean Spaces, so I migrated 12,108 records across 681 workflows to Cloudflare R2, keeping stored workflows connected to their media.
Cloud migrations. When our Azure credits ran out I moved us off Azure and retired 12 databases across 4 environments. Later, DigitalOcean’s uptime problems led me to move the backend to Google Cloud Run: both platforms ran in parallel until the DNS cutover, and I documented the deployment and migration procedure.
Other work at Mando the copilot, the marketplace, shared libraries, APIs and data safety
My first year went mostly into the copilot (it later became the product’s search) and the expert marketplace. For the copilot I built the answer interface from scratch: answers stream through the Vercel AI SDK and LangChain, inline citations are numbered and de-duplicated, and inference endpoints that had been public now need an API key.
For the marketplace I built a booking flow, expert chat and video calls on Azure Communication Services (the calls later moved to Twilio Video with recording), and Stripe Connect payouts to experts. Alongside it I reworked the Stripe subscription service for seat-based pricing behind customer plans, including seat changes and onboarding.
I extracted the shared UI and design-token libraries, and added the first frontend CI, Storybook tests and per-PR Storybook previews.
I built the expert-training and labelling pipeline: expert answers pass through a verification state machine, and an aggregator merges them into the labels that improve the copilot.
I ran an API quality programme: a CI coverage report found that lazy routers exposed only about 15 of roughly 925 routes to tooling, and I moved the generated client into a shared typed package that the frontend and the extension both use.
After a backfill brought deleted docs back in production, I standardised soft deletes behind one mixin, unifying 4 delete expressions and 5 read predicates.
I brought the dashboard’s load from over 10 seconds to within its 5-second target by decoupling the source fetch and caching with React Query.
SOC 2 preparation: an audit of the backend API, and short-lived tokens with refresh. Later I found a mass-assignment flaw that let organisation members write billing fields and wrote the fix: an allowlist of editable fields and a partial unique index.