Inbox & Telegram Triage
Classify, label, route, and draft responses. Escalate uncertain cases to a human on Telegram with one tap to approve or edit.
Private beta · Q4 2026
Meet Vega
The LLM automation platform for B2B operations.
Vega turns the repetitive work inside your company — inbox triage, Telegram support, data reconciliation, document extraction — into autonomous Python agents, deployed on your AWS account, with full audit trails and cost caps.
// The problem
Every B2B company has a layer of invisible work.
Ten to thirty percent of payroll in most operations teams goes to tasks that are deterministic, rule-driven, and infuriatingly repetitive: classifying a ticket, copying a value between two SaaS tools, answering the same Telegram question for the 400th time.
General-purpose LLMs are great at these tasks — in isolation. Turning them into something your CFO can trust in production is a completely different engineering problem. That is what Vega is built to solve.
// The solution
Vega = agent runtime + connectors + guardrails.
Define a workflow in plain language and one YAML file. Vega compiles it into a Python agent, wires it to your tools (Telegram, IMAP, Postgres, HubSpot, Stripe, Airtable…), runs it on your AWS account inside a hardened Lambda or Fargate service, and gives you one dashboard to see every run, every cost, every escalation.
When it is not sure, it asks — on Telegram, by default. No black boxes.
// What ships
Four capabilities that replace most ops work.
Document Extraction
Invoices, contracts, shipping docs — turned into typed JSON records and pushed straight to your ERP or data warehouse.
Cross-system Reconciliation
Nightly agents that compare Stripe, HubSpot, and your database, open a ticket for every mismatch, and resolve the simple ones automatically.
Observability & Cost Caps
Every run, every prompt, every token, every dollar — visible in one dashboard. Hard caps stop a runaway agent before it costs you anything scary.
// How it works
From plain-text brief to production agent in days.
-
01
Describe the workflow
Write the task in plain English. Upload a few examples of inputs and expected outputs. Commit a short YAML — or let Vega draft it.
-
02
Pick your connectors
Telegram, IMAP, Postgres, Stripe, HubSpot, Airtable, S3, custom HTTP — check the boxes you need. Secrets stay in your AWS Secrets Manager.
-
03
Vega compiles & deploys
A Python agent is generated, unit-tested against your examples, deployed as a Lambda or Fargate service in your own account, and wired to CloudWatch.
-
04
Supervise & iterate
Watch every run in the dashboard. Approve or edit uncertain cases from Telegram. Tune thresholds. Ship new versions without downtime.
// Technical highlights
Built for CTOs, not demos.
Your cloud, not ours
Vega deploys into your AWS account via a read-only Terraform module. Your data never leaves your VPC, and you can revoke access at any time.
Budgets, caps & HITL
Per-agent budgets, per-run token caps, and human-in-the-loop gates for any action classified as irreversible. Every decision is logged.
Claude now, others later
Today we default to Claude via AWS Bedrock. The runtime is model-agnostic — you can point any workflow at a self-hosted model when that makes sense.
// AWS Bedrock
Why Bedrock is the LLM control plane for Vega.
Every model call in every customer agent goes through AWS Bedrock. That is not a convenience — it is a deliberate architectural choice that unlocks four things we could not ship otherwise.
IAM-native security
Every InvokeModel call is authenticated by the Lambda execution role — no long-lived API keys, no secrets rotation, no leaked tokens on GitHub. This alone unblocks most enterprise pilots.
EU data residency
Bedrock in eu-central-1 lets us contractually guarantee that customer prompts and completions never leave European infrastructure — a hard requirement in every DPA we sign.
Bedrock Guardrails
Managed PII redaction, content moderation, and denied-topic filters — configured per agent from Terraform. Would take weeks to build; with Bedrock it is a resource block.
Knowledge Bases + Agents
Managed RAG over S3 and managed tool-use orchestration without operating our own vector store or agent loop — we focus on the workflow, AWS operates the infrastructure.
// Model choice
Claude today, model-agnostic always.
Vega routes each workflow to the right Claude tier on Bedrock: Claude Opus 4.7 for deep multi-step reasoning and agentic coding, Claude Sonnet 4.6 as the balanced default for most workflows, and Claude Haiku 4.5 for high-volume classification and cheap pre-filtering. Swapping tiers is a single line in the agent config.
Bedrock’s Converse API exposes every other major foundation-model provider under the same interface, which lets us match the workload to the model without rewriting code:
- Amazon — Nova Pro / Lite / Micro / Premier, Nova Canvas (image), Nova Reel (video), Titan (legacy)
- Meta — Llama 3.3 and Llama 4 (vision & reasoning)
- Mistral AI — Mistral Large 2, Mixtral, Pixtral (multimodal)
- DeepSeek — R1 reasoning models
- OpenAI — gpt-oss open-weight models for coding & scientific analysis
- Cohere — Command R+, Embed v3 (retrieval & embeddings)
- AI21 Labs — Jamba
- Stability AI — Stable Diffusion 3 (image generation)
// Cost & scale
Predictable economics, from pilot to production.
Bedrock's on-demand pricing is ideal for pilots; Provisioned Throughput kicks in once a workflow stabilises and the economics flip. We forecast and cap spend per-agent from day one, and every token is attributable in CloudWatch and CloudTrail.
Our internal target: under $0.02 per successful agent run at beta, trending to sub-cent at production scale.
Private beta is opening in Q4 2026.
Ten design partners, fixed discount for life, direct Slack line to engineering.