Case study
Real-time collaboration MVP
Shared control in the browser, shipped in four months
Oblong Inc.
- 0-1
- AI/ML
- Enterprise collaboration
- MVP
Summary
Role
Lead Product Manager
Timeline
Jan 2022 – Nov 2022
Team
Product, design, and engineering; LLM-assisted drafting for specs/exploration with human sign-off.
Domain
Enterprise collaboration; real-time multi-user interaction.
Metrics moved
MVP shipped in four months.
Key constraints
Short runway; enterprise security and UX expectations alongside build decisions.
Systems involved
React/TS + WebRTC in browser. Node signaling, WebSockets, Redis, Postgres, Coturn; AWS ECS/ALB/CloudFront; GitHub Actions. Applied AI on session data the product already captured.
Stack
Tech stack
Browser-first real-time stack, small observable backend, applied AI on session artifacts—not a bolt-on chatbot.
Real-time client
- React
- TypeScript
- WebRTC
Application services
- Node.js signaling
- WebSockets
- Redis (presence / ephemeral state)
- PostgreSQL
Infrastructure
- AWS ECS
- ALB
- CloudFront
- Coturn (STUN / TURN)
- GitHub Actions CI
Applied AI
Session-grounded assists with review paths.
- LLM-assisted PM workflows
- In-product ML on collaboration data
How the collaboration MVP shipped in four months
Simultaneous truth in the browser, bounded AI, stack a small team can run.
Charter and slice
Timeboxed discovery, then a thin vertical slice demoable every week.
WebRTC client + signaling spine
React/TS + Node signaling, Redis presence, Postgres metadata, Coturn through NAT.
Simultaneous truth filter
Features had to earn placement vs “shared control?”—else backlog.
Applied AI with receipts
Recaps, action items, search: preview/edit, triggers, eval hooks—not one big chat UI.
Enterprise hardening in parallel
Infosec and procurement answered alongside UX so pilots matched what eng would stand behind.
Overview
Led Oblong’s real-time collaboration MVP from charter to ship in ~four months. Headline: simultaneous multi-user control over WebRTC—not one presenter, passive audience. Applied AI embedded in real moments (summaries, action items, session search): narrow intents, preview/edit, measurement—not a generic chat surface.
Business context
Pre-pandemic Oblong sold enterprise multi-screen room installs plus proprietary screen share. Remote work broke the boardroom-centric model; team pivoted to remote-first collaboration from scratch.
Problem
Prove a new interaction model and product direction without a multi-year runway. Traditional roadmaps would have buried the riskiest assumptions under polish.
Constraints
Four months; small team; procurement and infosec in parallel with build. LLM tooling needed guardrails so speed did not become sloppiness.
Role & ownership
Lead PM: scope, sequencing, acceptance, customer narrative. Qualitative interviews prototype→beta; weekly product changes from feedback. AI: narrow intents, explicit I/O, human review, instrumentation—intelligence as workflow acceleration. With design/eng: crisp “simultaneous truth layer”; ML bounded to observable, testable behaviors.
Goals & metrics
Credible MVP on schedule; differentiated multi-user interaction in pilots; first ML enhancements (search, summarization, session intelligence) without destabilizing core sync.
Approach
Every feature candidate passed “does it require simultaneous truth?”—else backlog. Parallelized real-time spikes while UX iterated on paper.
Decisions & tradeoffs
Cut broader admin console for reliability/latency. Smaller testable AI scopes (preview/edit, eval hooks) vs one monolithic “AI surface”—bounded regression and trust risk in the MVP window.
Cross-functional leadership
Weekly slice reviews with cut/defer/prove; LLM-assisted spec/test drafting with human sign-off. Sales and solutions engineering in the loop so demos matched what eng would defend.
Execution
Two weeks customer calls + competitive teardowns, then a thin vertical slice demoable weekly. Stack: WebRTC in browser; React/TS clients; Node signaling/WebSockets; Redis presence; Postgres session metadata; Coturn STUN/TURN; ECS/ALB, CloudFront, GitHub Actions CI. Qualitative interviews through delivery. AI: small assists—structured recaps (Decisions / Actions / Open questions), action-item extraction with confirm/edit, auto-tagging + search—with triggers, preview/edit, measurement before scope creep.
Outcomes
Working MVP on the four-month timeline; pilots experienced shared control vs presenter-as-gatekeeper. ML reduced repetitive setup in demos and early deployments.
What changed
Prospects compared Oblong to more than screen share; conversations moved to workflow fit and data handling. Internal pattern for LLM-assisted delivery without skipping review.
Lessons learned
LLM value was prep and communication speed—not replacing scope judgment. MVP sold when the interaction model felt inevitable in five minutes; everything else was supporting evidence.