Analyst agent · Finance · Ops · Internal tools

Ask the data Get the SQL with the answer

We engineer auditable LLM agents over your SQL warehouses and internal documents—using LangGraph planner graphs, self-checking Text-to-SQL, typed MCP tools, and per-call Langfuse tracing so every answer comes with its SQL, sources, and cost.

Problem
Managers wait days for a report that one plain question should answer.
Proven
Text-to-SQL agent for stock analysis. Medical RAG platform with LLMOps tracing.
analyst.untern.ai / finance / thread-218
Why did gross margin drop in Q3? Which suppliers drove it?
Planned 3 sub-questions
Found 3 tables
Wrote SQL
Self-check: dry run passed, 1 column fixed
Ask a follow-up…

Architecture

Built for production Not a slide deck

Also in agents · Voice AI

Need a voice agent that knows when not to act?

Alongside Text-to-SQL and data agents, we build bilingual English and Vietnamese voice agents (like LifeMate) with hotword-biased speech recognition, an explicit act / ask / quiet decision policy, and typed, logged MCP tool calls.

What you get

Concrete outputs Measured on your data

01

Answers with citations

Answers over PostgreSQL, ClickHouse, PDFs, and wikis return the exact SQL queries and source rows in a Next.js / React chat workspace with an audit trail.

02

Self-correcting SQL

A LangGraph planner breaks questions into sub-queries, drafts SQL with few-shot templates, dry-runs and self-checks failures with a judge before answering.

03

Cost per question visible

Every LLM call and MCP tool execution is traced in Langfuse with per-call latency and cost, supporting Anthropic, OpenAI, Gemini, or self-hosted vLLM.

How we deliver

Small team Short loops

  1. 01

    Start from your real data

    A week with your documents, cameras or logs. We agree on one number that means “working” and freeze a test set before we train anything.

  2. 02

    Prototype behind a gate

    Baselines first, then models and agents. Every change runs against the frozen set, so progress is a chart, not an opinion.

  3. 03

    Harden for production

    Quantize, serve, trace and cost it. Retraining, rollback and monitoring are built in, not bolted on.

  4. 04

    Hand over, stay close

    Your team gets the code, the decision records and a walkthrough. We stay on call while it settles.

FAQ

Common questions Direct answers

How do you keep an analyst agent from hallucinating numbers?
The agent never invents figures from memory. A LangGraph planner generates SQL over PostgreSQL or ClickHouse using few-shot templates, runs a dry-run and judge self-check to fix broken queries, and displays the exact SQL and cited rows next to every answer.
Can we run self-hosted models instead of third-party LLM APIs?
Yes. We build against frontier APIs (Anthropic, OpenAI, Gemini) as well as self-hosted models served via vLLM when data privacy, latency, or unit cost require keeping inference inside your infrastructure.
How do you monitor agent reliability, crash recovery, and token cost?
Tools run through typed MCP interfaces and every step is traced in Langfuse so per-question cost and failure modes stay visible. We also benchmark framework resume behavior under hard process kills before standardizing on an orchestration layer.
Do you build voice agents in addition to chat and SQL agents?
Yes. Our voice agent architecture combines streaming 16 kHz English and Vietnamese speech recognition with hotword biasing and an explicit act / ask / quiet decision policy so spoken requests only trigger audited tool calls when intent is unambiguous.

Have a problem AI should solve?

Tell us what the data looks like and what “working” would mean for you. One of the founders will reply.