Live source: SUMMARY.md on GitHub · Auto-rebuilt on every streaming deploy-web.

Bitflix — Project Summary for the Bitmovin Team

A one-day, end-to-end OTT streaming demo built by Stefan + Claude Code on 2026-04-26, exercising the full Bitmovin stack (Encoding, Player, Analytics, AI Scene Analysis, Stream Lab, Live Encoder) on AWS — with two goals:

Build something that actually works, end-to-end, without cutting corners.
Document every place the Bitmovin product gets in its own way for an autonomous coding agent or first-time developer.

The output is a working demo at https://bitflix.slederer.com/ plus 37 detailed product/DX recommendations at https://bitflix.slederer.com/web/feedback.html.

GitHub: https://github.com/slederer/OS-streaming-v2

What got built

Live demo: https://bitflix.slederer.com/

Surface	What	Stack
Catalog	10 open-source films (Blender CC-BY + Public Domain) with hover-to-preview clips, posters built from the trick-play sprites	Static HTML/JS on CloudFront
Per-title VOD	H.264 + AAC, HLS + DASH, 3-rung Per-Title bitrate ladders, trick-play sprite + WebVTT, Bitmovin AI scenes.json with named characters/objects/dialogue, heuristic ad-marker overlays	Bitmovin Encoding Templates → S3 → CloudFront
Player	Bitmovin Player Web SDK v8 with chapter markers (white ticks) + ad-marker overlays (red ticks), Netflix-red theme over `bmpui-*` classes, custom Bitflix wordmark	Bitmovin Player + Analytics, custom CSS
Live (button-launched)	"▶ Start 10-min Live Session" provisions a fresh Bitmovin Live Encoding, ffmpeg pushes Big Buck Bunny on loop for 10 min, auto-cleanup, viewable on the same player	API Gateway → Lambda → Bitmovin Live + ffmpeg in Lambda
Stream Lab QA	20 countries × Chrome/Linux real device-matrix tests, 11/11 passing per country	Bitmovin Stream Lab targets API
Geo perf test	50 countries × real Chromium via residential proxies, time-to-playing 3.7-13.7s spread	browser-use Cloud (`wss://connect.browser-use.com?proxyCountryCode=<cc>`)
Feedback page	recommendations.md auto-rendered to `/web/feedback.html` on every deploy	Python `markdown` + custom Bitflix theme

Numbers

10 titles ingested + per-title-encoded in ~5 min wall-clock (parallel Bitmovin VOD encoding)
9/10 titles got rich Bitmovin AI Scene Analysis automatically (the implicit-by-account behavior is excellent — but undocumented, see #25)
20-country Stream Lab run: 11/11 success per country, mean startup 2,248 ms, but all from one Klagenfurt cluster — the "VPN location" knob is egress-only
50-country browser-use Cloud run: 48/50 success, median 6,730 ms, range 3,758 ms (Sweden) → 13,672 ms (NZ) — ~100× the variance of the Stream Lab run, because real residential proxies expose actual CDN-edge routing
Live encoding: 5-min and 30-min sessions both ran end-to-end, with 20 manifest samples returning 200 OK, TTFB 333-623 ms
End-to-end build time: ~10 hours total across 5 working sessions
Recurring AWS cost: ≈ $1-5/mo (S3 storage + CloudFront + Lambda idle = ~$0)
Bitmovin spend on this build: ~$15 encoding + ~$5 live + ~$2 Stream Lab tests = ~$22 total

Architecture (one diagram)

                ┌───────────────────────────────────────────────────────────────┐
                │                       Bitflix demo                            │
                └───────────────────────────────────────────────────────────────┘

          ingest CLI                        Encoding Templates
   archive.org ──────▶ S3 input ─────▶ Bitmovin Encoding ──▶ S3 output ─▶ CloudFront ─▶ bitflix.slederer.com
                                          │ Per-Title                       │
                                          │ Sprites                         │ Cloudflare DNS
                                          │ AI Scene Analysis (implicit)    │
                                          ▼                                  │
                                    Default-Manifest API                     │
                                          │                                  │
                                          ▼                                  ▼
                                    master.m3u8 + stream.mpd       ┌────────────────┐
                                                                    │  Browser       │
                                                                    │  Bitmovin      │
                                                                    │  Player + UI   │
                                                                    │  Analytics SDK │
                                                                    └────────────────┘

          "▶ Start 10-min Live"
                Click ──▶ API Gateway ─▶ live-start Lambda (returns URL <30s)
                                                    │ async
                                                    ▼
                                            live-push Lambda (≤15min, ffmpeg layer)
                                                    │
                                                    ├── boto3 GetObject (S3 input → /tmp)
                                                    └── ffmpeg → RTMP → Bitmovin Live Encoder
                                                                                 │
                                                                                 ▼
                                                                       same S3 → CloudFront → Player

          QA / Verification
                ├── pytest (unit tests, 14)
                ├── streaming verify (curl smoke)
                ├── streaming streamlab-test (Bitmovin Stream Lab, 20 countries × Chrome)
                ├── streaming geo-browser (browser-use Cloud, 50 countries × real Chromium)
                └── streaming browsertest (local Playwright E2E for player+analytics)

What was hard, and why it matters

This was the CEO of Bitmovin building a demo on Bitmovin. We hit 37 distinct friction items, almost none of which were missing features. They were:

Surface-area mismatches — playerCfg.ui = true silently breaks the UI but the player "loads"; segment_{number}.m4s is treated as a literal filename in templates; gb rejected for uk in browser-use
Discoverability gaps — Stream Lab device-matrix testing requires a 4-segment URL path (/streams/<sid>/targets/<stid>/jobs) that an LLM agent will not guess; I wrongly concluded the API didn't exist before being corrected
Doc/SDK lag — AI Scene Analysis is heavily marketed but absent from bitmovin-api-sdk-python 1.264.0; the JSON schema for Encoding Templates lives in a different repo from the docs page; bitmovin-api-sdk versions ≤1.234 pin requests<2.25.0 (incompatible with anything modern)
Dashboard ↔ API divergence — multi-license Analytics dashboards need explicit license selection (we had two, picked the wrong one, "no impressions" panic ensued); Stream Lab notified is a terminal status the API uses but the docs don't list; testresults expose durationMs per test but no summary.metrics.{startupTime,avgBitrate} rollup

The full list, with reproductions, error messages, and one-line fixes per item, is at https://bitflix.slederer.com/web/feedback.html (also recommendations.md at the repo root).

Top 5 fixes that would 10× the developer/agent experience

Ranked by adoption impact:

playerCfg.ui should reject true at construction. Silent UI failures are unforgivable. (Item #17)
POST /encoding/templates/validate dry-run endpoint. 5 of my 7 wasted encoding cycles would not have happened. (Item original-1)
Multi-license Analytics dashboard default = "All licenses" + SDK is_default flag. Eliminates the "no data showing" support tickets entirely. (Item #18)
Make AI Scene Analysis explicit. Today it runs implicitly at the account level when enabled, with no opt-out and no lifecycle messages. Surface it. (Items #25, #26)
Structured-error responses with path + expected + suggestion. "An unexpected error occurred — please contact support" is the worst possible thing to return to an automated client. (Item original-2)

Insights worth product-team discussion

Topic	What we learned
Per-Title encoding	Excellent. Drop in `mode: PER_TITLE_TEMPLATE` + `start.perTitle.h264Configuration.autoRepresentations` and you get sensible bitrate ladders with no manual tuning. Showcase feature.
AI Scene Analysis	Output quality (when it runs) is genuinely impressive — character names, object categories, dialogue, scene descriptions. 9 of 10 titles got real AI output without me asking for it (account-level implicit). Failure for 1 title (Spring) traced to source bitrate (~800 kbps for 1146×480 — Bitmovin AI seems to need a quality floor).
Stream Lab — what it is	World-class for device-matrix QA: 31 real device targets including LG WebOS 2016-2024, Samsung Tizen 2016-2024, Vizio, Xbox Series S, PlayStation 5. Nothing else in the market touches this.
Stream Lab — what it isn't	A geographic CDN-performance benchmarker. The "VPN location" knob simulates egress only; the Chromium runner stays in `eu:klu:b2`. Across 20 countries, startup-time spread was 140 ms — ~100× compressed vs reality. We added a comparison table to recommendations.md (item #36).
Live Encoder	Provisioning + RTMP push + manifest serving + cleanup all worked end-to-end. The `live.get(encoding_id)` returns errorCode 2023 ("not available!") for 60-180s after `live.start()` — that 2-3 min provisioning gap is the single thing that forces a multi-Lambda async pattern when wiring up a browser-callable trigger. A `wait_for_rtmp_ready(timeout)` SDK helper would make this a 1-call setup.
Player + Analytics	Integrated in ~10 lines of JS. The `ui: true` silent UI break (item #17) is the only sharp edge. Worth fixing — automated tests can pass while the UI is fully broken.

What this build proves about Bitmovin's market position

The product is genuinely capable. None of our 37 items are missing-feature gaps; every one is polish. That's a strong position to start from — the alternative would be missing capabilities, which is much harder to fix.

The specific differentiator that emerged: Bitmovin's per-title encoding output quality + sprite trick-play + AI Scene Analysis output quality, taken together, are visibly better than the Mux / AWS Elemental MediaConvert / JW Player stacks. We didn't have to manually tune anything to get a presentable per-title ladder, sensible scene chapters with dialogue summaries, and clean trick-play scrubbing.

The specific gap that emerged: there's no "single canonical worked example" anywhere in the product surface. The closest is bitmovin-api-sdk-examples, but it's imperative SDK Python — the wrong audience for someone authoring an Encoding Template YAML. A "starter pack" repo with one canonical YAML per common workflow (per-title VOD, sprites, AI, multi-DRM, live) would dramatically reduce the time-to-first-success that this build is, in a sense, a demonstration of.

What's next (suggestions)

If anyone on the team wants to extend this:

Switch CloudFront to PriceClass_All and re-run the 50-country browser-use test. Should collapse the APAC/AF tail (currently 10-14s) to ~5s. Free win on Bitflix's side; great content for "your CDN edge selection matters" demos.
Add multi-DRM: Widevine + FairPlay + PlayReady via the Bitmovin Encoding's DRM block + DoveRunner. Then re-run Stream Lab against the smart-TV / console targets — that's where Stream Lab is uniquely strong.
Add MediaTailor SSAI: insert real ads at the AI-detected ad-marker positions in the player.
Schedule a one-time agent in 4 weeks to re-run the same 37-item probe against any fixes that ship, and diff the friction list. (/schedule is set up for this.)

Reading order for the team

This document.
https://bitflix.slederer.com/ — see it work. Click "▶ Start 10-min Live Session".
https://bitflix.slederer.com/web/feedback.html — the 37 items. Skim P0/P1, skip P2 unless triaging.
https://github.com/slederer/OS-streaming-v2 — the code. templates/per_title.yaml.j2 and infra/lambda/live_push/handler.py are the most "this is how you actually do this" parts.

— Stefan Lederer, 2026-04-26.