Skip to content
All works

2026ShippedProduct

Feige Outbound Platform

feige-outbound-platform

A local pipeline that connects calling, recordings, transcripts and follow-up without a human babysitter.

Role
Design & development, solo
Year
2026
Status
Shipped
Stack
TypeScriptFastifySQLiteMiniMax ASR

Outbound calling is where a salesperson’s words turn into data — and where most of that data gets lost. The Feige console lets you read call logs, listen to recordings and dial out, but recordings expire, transcripts are incomplete, and there is no API and no callback. Anything a script or an agent might do with a call needs a human keeping a browser open.

The problem

I split outbound work into three jobs: dial, keep the call log and recording, turn the recording into readable text and analysis. Upstream only offers a web page for all three. Two practical problems make it worse: the login expires after seven days and the Aliyun slider is hard to automate; call history is scattered across pages, and recordings disappear after a window.

What I built

FOP is a resident local platform. Northbound it has four entry points: REST v1 (Fastify, loopback plus bearer, OpenAPI 3.1 and Swagger), a CLI, an MCP stdio server and a SKILL document. Southbound it automates the page: multi-account login including the slider, seven-day JWT keep-alive, incremental and full call-log sync, and outbound dialing with a queued state machine.

The most important part is that recordings stop disappearing. They are downloaded during sync, hashed and archived as WAV files by year/month/day. Then transcription happens 100% locally.

Architecture notes

  • Local ASR: MiniMax asr-1.0. The recordings turn out to be 8 kHz stereo with the agent and the customer on separate channels, so splitting the channels gives speaker separation for free — far more reliable than diarization guesses. Long calls are cut at silence points; chunk results are cached so a retry resumes instead of restarting. A term dictionary normalises the nine common mishearings of the company name.
  • AI post-processing: DeepSeek with a fixed prompt and a JSON schema; validation failures escalate through a retry chain. The output is a structured summary, intent, keywords, action items, risk flags and quality notes.
  • Storage and runtime: SQLite WAL with 15 tables, WAV archive, SSE events and a recoverable job queue. systemd keeps it running; Xvfb handles headless login.

What I learned

93/93 unit tests and 27/27 acceptance checks pass. Writing limits honestly matters too: upstream has no SIP line, so real dialing is callback mode and the agent’s phone must answer first; v1 does no AI voice calling; ASR concurrency on a default account is one, which makes full backfills slow. All of that is in the README.

The bigger takeaway is data ownership: when recordings and transcripts live on your own machine, the sales process becomes something you can query and review for the first time.

93 / 93

Tests passing

27 / 27

Acceptance checks

35

REST operations

Next projectTennis-Ball Robottennis-ball-robotView project