Why We Stopped Watching Your Whole VOD (and What We're Trying Instead)

UpForge is operating normally again. If your desktop Valorant analysis failed recently, or the feedback felt vague, you deserve a straight explanation — not marketing fluff.

This post covers what was going wrong, what we changed, what you should see in a report now, and what we still don't know.

---

The old approach (and why it fell short)

For desktop recordings, we used to send your entire match VOD through video review at about 1 frame per second.

At that frame rate, the model can follow round flow and spot obvious positioning mistakes. It struggles with duels — the moments that actually decide ranked games:

• Did you wide-swing or jiggle?
• Was your crosshair at head height or on body/legs?
• Did you commit before your teammate traded?
• What happened in the two seconds before you died?

When those details are blurry, coaching defaults to generic advice. "Work on crosshair placement" and "don't over-peek" are true, but not useful if you can't tie them to a specific round.

---

The new approach: death moments

Desktop Valorant coaching now uses a moment-based pipeline:

• Short clips around your deaths — the desktop app (or our backfill from your VOD) extracts duel windows for each death we can identify.
• Higher frame rate on those clips — up to 24 FPS on the moments that matter, instead of 1 FPS across the whole match.
• Riot match data — rounds, kills/deaths, economy, and callouts where your recording captured them.
• Coaching synthesis — feedback tied to specific rounds and timestamps, e.g. "Round 7 @ B Main" rather than blanket tips.

The goal is simple: tell you what went wrong in the fights that cost you, in language you can act on next game.

Note: This applies to desktop Valorant recordings. Web/YouTube VOD analysis still uses a different path. We're focused on getting desktop coaching right first.

---

What you should see in a report

When the pipeline runs cleanly, expect:

• Round-linked improvements — each major point references a round or moment, not just general habits.
• Untraded death callouts — when you died without a trade, the report should say so and why it mattered.
• Duel detail when the clip supports it — peek type, crosshair height, commit timing. If the clip is unclear, a good report admits uncertainty instead of guessing.
• Category breakdown — aim, positioning, utility, etc., grounded in what was observed.

We're not claiming every report hits this bar yet. Some will. Some won't. That's the honest state of things.

---

What broke (and what we fixed)

Some analyses failed outright even after we shipped the new approach. Root causes included:

• Duel clips missing from storage — the pipeline fell back to low-FPS full-VOD review and produced thin signal.
• A callback bug — duel observations weren't always passed back to complete the job, so reports failed integrity checks with "insufficient duel observations."
• Stale job state — re-queued analyses could be killed immediately if old timestamps weren't cleared.

We've fixed these on our end. If you had a failed analysis in the last few weeks, it's worth trying again — or open a support ticket and we can rerun where possible.

---

What we're still learning

This is early work. We believe death-moment review is the right direction, but we don't yet know how good it can get.

Open problems we're actively working on:

• Clip quality — not every death window is framed perfectly; uploads can be incomplete (we can sometimes backfill from the full VOD).
• Vision misses — the model can misread a fight or skip detail when lighting, UI clutter, or chaos gets in the way.
• Feedback tone — advice that helps an Immortal player can feel wrong for Gold; we're tuning rank-aware language.
• Constant experimentation — clip length, review steps, prompt structure, and how we combine video with match stats.

We run internal fixture tests (same match, different clip sources) to compare before we ship changes. Your feedback and support tickets directly shape what we fix next.

---

How you can help

If you run a desktop analysis:

• Note what felt specific vs generic.
• Tell us if a death was misread or advice felt wrong for your rank or agent.
• Failed analysis? Submit again from the desktop app, or open a ticket in our Discord (#ticket-support).

We're a small team building in public. Useful coaching matters more to us than flashy feature lists.

Thanks for sticking with us.

— Adam, UpForge