NewAsk by voiceSpeech-to-text byAssemblyAISpeak now
SeravionPESTPILOT
AI Coworker for Seravion Connect

Ask your operationscustomersfinanceinventoryHRtrackingticketsanalyticsoperations
anything, rapid.

Jobs, technicians, contracts, invoices, stock. Ask in plain English and get the answer in seconds, with the exact records behind it.

Try it nowSee it answer a question
Read-only by designLive ERP dataExcel & PDF export
What is the support tickets distribution?
Here's the tickets count by branch, open against resolved
XLSXPDFDetails

Runs on the infrastructure and models you'd choose yourself

NeonPostgreSQLHetzner CloudGoogle CloudRenderGroqHugging FaceOpenRouterKaggleAssemblyAILangChainFastAPINext.jsMapbox
The gap

Your ERP speaks SQL. Your team doesn't.

The answer is already in your database. Seravion translates the question, finds the right tables, writes and checks the query, in the open, step by step.

01 · YOU ASK

Semantic search across 213 tables
support_ticketsbranchescustomersamc_contractsjobsemployeesinvoicestechniciansstock_itemspetty_cashattendancevisits
02 · SERAVION WRITES
ParsedSchema-checkedDry-runRead-only
03 · YOU GET
Nothing runs until the query passes every check.
Reading the questionSample data · read-only connection
How it works

An analyst that shows its work.

01

Understands the question

Semantic search across every table, column and relationship finds where the answer lives. No keyword matching, no LLM guesswork for routing.

support_tickets0.91
branches0.84
ticket_logs0.42
02

Thinks twice when it's unsure

03

Writes and checks the query

04

Answers, and shows its work

Not every answer is a table.

Seravion picks the right shape for what you asked, automatically.

KPI

"How many?" gets a number, not a spreadsheet.

Chart

Trends and comparisons, drawn for you, with the data one tap away.

Table

Operational lists you can scan, page through and export.

Cards

People, customers and branches as records you can open.

Every format exports to Excel and branded PDF, from the exact rows you're looking at.

One chat for
operations.

Leaner by design.

We rebuilt the engine so routine questions cost less and wait on nothing they don't need.

−0%¹prompt tokens per question¹ Same 6 questions, same provider, old vs new pipeline (55,158 → 12,016 prompt tokens).
0²LLM calls to route 73% of questions² Routing uses embedding search; the decision model runs only on the ~27% of questions retrieval is unsure about (measured on 90 real questions).
< 1 ms³to search the whole schema³ Exact vector search over 213 table embeddings.
2 → 1⁴query executions per answer⁴ Each question's query now runs exactly once (it used to run again to build the table); the same result feeds the answer, the UI and exports.
0+⁵tables understood, 290+ relationships mapped⁵ Per tenant schema, from the live catalog.

Read-only. Validated. Yours.

Read-only, twice over

Queries run in read-only transactions with a timeout and row cap; write statements are rejected before they run.

Validated before it runs

Every query is parsed, schema-checked and dry-run first. One repair attempt, then an honest "couldn't pull that up", never a made-up number.

Honest when it's empty

No data means "nothing found", not a confident guess.

Secrets in a vault

API keys and credentials live in Google Secret Manager, not in config.

Tenant-scoped

Each company's questions run against its own schema.

Which branch has the most support tickets?

Mumbai leads with 142 tickets, 20% more than Pune.

Details
Routing
Decision
Tables retrieved
SQL validation
Repairs
Execution
Grounding
Citations
“142”←…
“Pune”←…

Serious infrastructure, boring in the best way.

DataPostgreSQL 15 on Hetzner Cloud · PostgreSQL 18 on AWS RDS
ComputeGoogle Cloud Run · Cloud Build · Artifact Registry · Secret Manager
RetrievalBAAI bge-small-en-v1.5 embeddings, in-process · NumPy exact vector index · pgvector-ready
DecisionTypeSafe Jev: structured routing + presentation decisions
ModelsGroq · OpenRouter · Hugging Face: Llama 3.x and gpt-oss-120b, multi-key rotation
Safetysqlglot parsing · schema catalog checks · Postgres EXPLAIN dry-runs
BackendPython · FastAPI · LangChain · SQLAlchemy · Upstash Redis plan cache
FrontendNext.js 14 · React 18 · Tailwind CSS · Recharts
ReportsWeasyPrint branded PDF · openpyxl Excel

Built for every desk.

“Which branch grew the most this quarter?”

Revenue, receivables and branch performance without a report request.

AI Coworkergroq · 3.4s · 2,108 tokens · 2 calls10:42 AM

Pune grew the most: revenue is up 24% on last quarter, ahead of Mumbai at 18% and Bengaluru at 11%.

Revenue Growth by Branch (%)
+0%
+0%
+0%
+0%
+0%
PuneMumbaiBengaluruDelhiChennai
5 records · sales_invoicesView dataXLSXPDF

Questions, answered.

No. It can only read, and the database transaction enforces that, not just the AI.

Stop waiting on reports.

Ask your first question in minutes.

Book a demo Talk to us
PestPilot chat in a browser: an overdue tasks trend question answered with a line chart