obsei

announcement

Obsei is back: private Voice of Customer for the agent era

Why we are reviving obsei after years of dormancy, what changed since our RBI HARBINGER recognition, and why customer feedback analytics now has to be private, open and agent-ready.

By 7 min read

obsei logo with the words: obsei is back, private Voice of Customer for the agent era

Obsei went quiet for almost three years. Today it is back, rebuilt from scratch, and it will stay fully open source.

This post is the story of how we got here, why I think Voice of Customer matters even more in the age of AI agents, and where obsei is going next.

Where obsei started

Obsei began in December 2020 as a small open-source Python library (first release on PyPI). The idea was simple: pull customer feedback from wherever it lives (app stores, Twitter, Reddit, support desks), run it through AI models for sentiment and classification, and send the result wherever a team works. That was before ChatGPT, when “AI” for most teams still meant training your own classifier.

Obsei is also how I learned AI. I started by contributing to other open-source projects, above all Haystack, and sending a few patches to Hugging Face libraries such as Transformers. Obsei grew out of that: transformers for sentiment and zero-shot classification, then small models I trained myself for the RBI hackathon, then early retrieval-augmented generation, all before most people had heard of a chatbot.

In June 2022 obsei was named a joint runner-up in the Reserve Bank of India’s first global hackathon, HARBINGER 2021, under the problem statement “Social media analysis and monitoring tool for detection of digital payment fraud and disruption”. The jury picked 24 finalists from 363 proposals from India and 22 other countries. RBI described obsei as an AI-powered, real-time social media analytics tool that understands customer feedback in their native languages.

That was the validation we had been looking for. Later in 2022, Girish Patel and I founded Oraika Technologies to turn obsei into a company.

From open-source contributions to a privacy-first VoC platform: obsei, 2020 to 2026.

Why we stopped

Recognition did not turn into a business. After the award we met several leading Indian banks and payments institutions. The conversations were encouraging, but we could not close a deal: we were engineers with no experience of enterprise sales, procurement or long security reviews. It turns out tech people find selling hard :). We also tried the entertainment industry and a lot of cold outreach. None of it stuck. Looking back, obsei taught me far more than AI: it was a full course in how to fail at sales, how to fail at marketing and how to fail at fundraising. Every one of those lessons shaped how obsei 1.0 is built and run.

We closed Oraika and moved on. I now lead technology at Bluepencil and am building Tashi, an end-to-end SaaS for the social sector.

I have a few open positions in my team at Bluepencil: building meaningful technology for the social sector and for privacy under India’s DPDP Act. If that sounds like you, message me on LinkedIn.

Why now

Even while nobody was working on it, obsei kept being used. The repository has more than 1,400 stars and 170 forks, and the old 0.0.x package is still downloaded from PyPI every month. People still need this, and an unmaintained library that copies raw customer text around is not what they should be running in 2026.

The problem never went away. If anything, it got bigger:

  • Customers talk everywhere. Reviews, tickets, calls, communities and surveys, in dozens of languages. No team reads it all.
  • Feedback is personal data. Names, emails, phone numbers, order ids and national IDs sit inside almost every message. Copying all of it into yet another SaaS, or pasting it into a public chatbot, is a privacy risk most teams can no longer accept.
  • Agents need ground truth. AI agents now write code, triage tickets and plan roadmaps. They should answer “what are customers saying?” from real, cited evidence, not from a guess.
  • Models finally fit on your own machines. Small open models, and now decision models that answer in a fraction of a second on a CPU, make private analysis practical for any team.

So obsei 1.0 is not a polish of the old library. It is a new product built around one rule: understand customers without moving their personal data anywhere it shouldn’t go.

Same mission, new rules: what changed between the first obsei and 1.0.

What obsei 1.0 is

  • Private by design. Personal data is redacted at ingest (national IDs for seven regions, checksum-validated where the format allows; names optional), authors become pseudonyms, the database is encrypted, and nothing leaves your network unless you allow it (privacy controls). Air-gapped is the default.
  • Bring your own sources and models. 20+ sources, from App Store to Zendesk to SQL, plus plugins. Any OpenAI-compatible model, from Ollama to Azure, and decision models such as Julia-1 for fast labels with real probabilities.
  • Themes across languages. The same issue raised in English, Japanese and Portuguese becomes one theme, with trends, near-duplicate detection and k-anonymous views: groups from fewer than k people are hidden.
  • Built for agents. A read-only MCP server lets Claude, Cursor, VS Code and your own agents query feedback with redacted quotes and record ids, never identities.
  • Act on it. Route urgent bugs to Jira, billing to Slack and unsure cases to a human, based on model confidence (routing).

Why you might use it

A few situations where obsei earns its place. Each one links to a complete, tested example configuration.

You ship an app and a release goes wrong. Reviews start piling up in six App Store countries and Google Play, in five languages. Obsei groups “can’t log in after the update” in English, Spanish and Japanese into one rising theme, and posts it to Slack before the support queue explodes. (app reviews example)

You run support and every ticket lands in one queue. A small decision model on a CPU reads each Zendesk or Freshdesk ticket once and answers typed questions: which team, how urgent, is the customer angry, do they want a refund. Urgent bugs go to Jira and on-call, billing goes to the billing channel, and anything the model is unsure about goes to a human. (routing example)

route:
  - review: [review]                     # the model is unsure: a person decides
  - when: {classify.intent: {is: bug, min_confidence: 0.8}, classify.fields.urgency: {gte: today}}
    sinks: [jira, oncall]
  - default: [lake]                      # everything else: your warehouse

You work in a regulated industry. Banks, insurers, health and public services often cannot send customer text to a third-party SaaS. Obsei redacts Aadhaar, PAN, CPF, IBANs, card numbers and more at ingest, encrypts the store, and runs air-gapped with your own model. This is the setting obsei grew up in: spotting payment-fraud chatter on social media for the RBI hackathon. (air-gapped example)

You serve several markets. One pipeline per region, each with its own intents, channel and privacy rules, while themes still show the issues they share. (India, Brazil and Japan example)

You build or use AI agents. Point Claude, Cursor or your own agent at obsei’s MCP server and ask “Why are Japanese users unhappy this week?”. The answer comes back with redacted quotes and record ids you can check, and never with a customer’s identity. (MCP guide)

claude mcp add obsei -- obsei mcp

You run a nonprofit or social program. Surveys, field reports and helpline messages are feedback too, often from vulnerable people. Obsei lets a small team see what beneficiaries are saying, in their own language, without exposing who said it. (surveys example)

You maintain an open-source project. Watch GitHub issues, Hacker News, Reddit and Bluesky for what users actually struggle with, and turn recurring themes into issues. (social listening example)

All examples are on the Examples page.

You can try the live demo or run it locally in five minutes with the quickstart.

Open source, for real

Obsei is Apache-2.0 and will stay that way. There is no company behind it, no hosted upsell and no open-core split. I will work on it in my free time, in the open.

The commitments obsei is built on.

The goal is for obsei to outgrow me. I want it to become a project worth donating to a neutral foundation such as LF AI & Data or the Apache Software Foundation (through its Incubator), so that the community, not a single vendor, decides its future. That needs contributors, users and honest feedback (ironic for a feedback tool, I know). The groundwork is already in place: an OpenSSF Scorecard, container images signed with cosign, a security policy and a contributing guide with DCO sign-off.

What’s next

obsei 1.0 is coming, and its release candidates are out:

uv tool install "obsei[mcp]>=1.0.0rc1"

Or run it with Docker, in an empty folder:

export OBSEI_DB_KEY="$(openssl rand -hex 24)" OBSEI_PSEUDONYM_SALT="$(openssl rand -hex 24)"
alias obsei='docker run --rm -v "$PWD:/data" -w /data --user "$(id -u):$(id -g)" \
  -e OBSEI_DB_KEY -e OBSEI_PSEUDONYM_SALT ghcr.io/obsei/obsei:1.0.0-rc.1'
obsei init && obsei run      # sample feedback in ten languages, redacted and stored encrypted

1.0 follows once real-world connector checks pass (see the roadmap).

If customer feedback matters to your team, and privacy matters too, give obsei a try, open an issue, or start a discussion. Obsei is back, and this time it is built to last.

Lalit Pagaria · LinkedIn · GitHub