Internal demo · May 5, 2026

Studio One PM Assistant

Describe a new project in plain English, Russian or Armenian.
Get the best-fit developer in 30 seconds, with cited evidence from real GitLab + Jira data.

🇦🇲 Հայերեն 🇷🇺 Русский 🇬🇧 English 🎙 Voice messages

Open in Telegram

@st1_assistant_bot

Scan to try it now

The problem

A new project lands. The first 20 minutes go to the same questions every time.

→ "Who's free for the next 6 weeks?"
→ "Who actually knows TypiCMS / Magento / Flutter?"
→ "Who shipped something like this last time?"
→ "Is anyone on vacation when this would launch?"

That's per-project per-PM, every week. We're going to compress it to 30 seconds, on our own infrastructure, with answers backed by real commit and ticket data — never made up.

How it works

Three steps. No magic.

PM describes the project

In Telegram, in any language. Stack, scope, integrations, deadline. Voice messages work too.

LLM reads team data

Qwen 2.5 Coder runs on our VPS. It reads team.yaml — a structured dump of every dev's GitLab commits, Jira history, current load, and skills.

Ranked recommendation

Top-3 with cited evidence — specific tickets they shipped, current load, why others were excluded.

What it answers today

Real PM queries, real replies.

Routing · EN

"Need to extend the Hall Builder venue editor with a drag-and-drop sections feature. React + Konva canvas. 5 weeks."

→ Mnatsakan top (cites TP-124 / TP-118 canvas tickets), Alik runner-up

Routing · RU

"Нужно собрать многоязычный TypiCMS-сайт для фонда — AM/RU/EN. 8 недель."

→ Erik or Alik on top, Gor flagged 100% loaded — reply in Russian

Honest no-match

"Need a native iOS Swift engineer for a banking app, 8 weeks."

→ "No one has documented native iOS / Swift experience." — won't manufacture a match

Vague input

"Client wants a website."

→ Asks one specific clarifying question — never guesses

Design choices

Why we built it this way.

On our infra

Qwen 2.5 Coder runs on a Studio One VPS. No project description, no team data, nothing leaves Studio One — ever. The model is open-source.

Honest by default

The system prompt is tuned to refuse when the data doesn't support a recommendation. If the team has no Salesforce experience, the bot says so — it doesn't pick the closest PHP person and call them an expert.

Learns from PMs

When voice transcription mishears a name, the bot offers alternatives. Tap the right one — it remembers, across users and across restarts. The team teaches the bot one correction at a time.

Real data only

Every skill, every "recent_work" line comes from a real GitLab commit or Jira ticket. The ingestion job re-runs nightly. No hand-maintained spreadsheets, no stale data, no opinions — just the audit trail.

Roadmap

What's next.

TODAY

Phase 1 + 2

Recommendation flow with cited evidence · trial period analytics

+ 4 wks

Phase 3

Auto-create Jira epic + skeleton tasks once the PM confirms a developer

+ 6 wks

Phase 4

Battle-tested Docker stacks for Magento 2 / Laravel / TypiCMS — bootable on a fresh VPS in 15 min

+ 8 wks

Phase 5

Auto-provision a project's full stack on a VPS · report URL + credentials back to the PM

The ask

Three PMs to use it 5 times each over the trial.

Pick someone curious, not someone we have to convince. We log every recommendation and every PM choice — at the end of the trial we'll see how often the bot agreed with the human, and that tells us if it's actually useful.

Try it now → @st1_assistant_bot

Built by Edgar Grigoryan · Studio One · May 2026