All posts
AIEngineering

Meet Anuradha: The HR Avatar Who Never Sleeps (And Never Says "Let Me Get Back to You")

AI Avatar

Aditya Chandrabhan Chaudhari
Aditya Chandrabhan Chaudhari
Admin
May 14, 2026
21 views 5 min read

The Problem

If you've worked at ConveGenius for more than a week, you know the drill. You want to plan a Diwali trip, so you Slack HR: "how many casual leaves can I take at once, and do they carry over?" You're flying to a State meeting in Shimla and you Slack HR: "is the cab from the airport covered under Local Conveyance or do I claim it under Domestic Travel?" You're about to refer a friend and you Slack HR: "what's the referral bonus on a senior hire, and when do I actually get paid?"

Multiply that by 500+ ConveGeniuses and one HR head, and you get the universal corporate truth:

The policy is written. The PDF is uploaded on Keka. The answer is somewhere. Good luck.

And the questions are real. Across our 16 policies, the day-to-day asks aren't exotic - they're things like:

  • Leaves - "Do I have nine casual leaves or twelve? What about menstrual leave? Do unused ones lapse on December 31?"

  • Local Conveyance - "Is my Uber to a client meeting reimbursable, and do I need a receipt under ₹500?"

  • Maternity Benefit - "How many weeks of paid leave, and can I extend with unpaid leave after?"

  • POSH - "I want to raise something, but quietly first - who exactly do I go to?"

  • Referral - "If my referral joins, when does the bonus hit my account?"

  • Guest House - "I'm going to the Gujarat office next month - am I eligible for the guest house or do I book a hotel?"

  • Asset Policy - "My laptop is two years old, when am I due for a refresh?"

  • Whistleblower - "Is what I want to flag covered under this? Can I stay anonymous?"

  • Transfer / Deputation - "If I move to the Delhi office for six months, what changes about my CTC and HRA?"

  • PIP - "If I'm on a Performance Improvement Plan, what does the next 60 days look like for me?"

  • Code of Conduct & Social Media - "Can I post about a feature I'm working on, or do I need approval?"

And here's the thing: Anuradha is genuinely approachable - door always open, Slack always on, never a "circle back" - which is exactly why she ends up answering the same five questions fifty times a week, on top of doing actual HR.

Every one of those questions has a crisp, written answer sitting in a policy PDF. The information isn't missing - it's just buried in a Drive folder nobody opens. So the answer-of-last-resort becomes "ping HR" - and HR ends up being a human FAQ bot instead of doing the actual work of HR.

It's not HR's fault, and it's not the employees' fault. It's a discoverability problem dressed up as a people problem.

So I asked myself a simple question: what if every one of those policies could just talk back?


The Idea: Give Anuradha her clone!

Not a creepy clone. A helpful one.

The pitch was: build an avatar version of our HR head - same face, same tone, same answers - that any employee can talk to, any time, in whatever mode feels natural. Voice if you're walking. Video if you want to feel like you're actually in a meeting. Text if you're in a stand-up and don't want to be caught chatting.

Three rules I gave myself:

  1. It must feel like talking to Anuradha, not to ChatGPT pretending to be HR.

  2. It must be accurate - wrong answers about leave policy are worse than no answers.

  3. It must give HR back their calendar.


The Solution: Anuradha's Avatar

Meet Anuradha - our AI HR companion!

What it does:

  • Video mode - a real-time avatar that lip-syncs and responds, perfect for "I want this to feel like a meeting" moments.

  • Voice mode - talk to her hands-free, she talks back. Great for grievances you'd rather not type out.

  • Text mode - fast, async, and you can attach your own files (offer letter, leave application, a screenshot of a payslip) and ask questions grounded in that document.

  • Policy downloads - when you ask "share the domestic travel policy," she actually shares the PDF. Not a summary. The real thing.

  • Usage insights - HR gets a dashboard of every conversation, every concern raised, every policy that gets asked about most. So instead of answering the same question fifty times, HR can spot patterns and act on them.

  • What's under the hood (the high-level version, because this isn't an engineering blog):

    • OpenAI's GPT-4o for the brain

    • Anam AI for the real-time avatar and voice cloning (yes, it really is her voice)

    • Cartesia for the voice-mode TTS

    • Zep for memory - Anuradha remembers your last conversation so you don't have to re-explain context

    • FastAPI + a sprinkle of vanilla JS for everything that glues it together

    • All 16 HR policy PDFs indexed locally, searched with BM25, so answers are grounded in the actual documents - not made-up

    The whole thing runs on a single Render instance. The most exotic infra decision was "let's put it in Oregon."


    The Impact (a.k.a. Why This Matters)

    This wasn't built as a tech demo. It was built because:

    • HR's calendar is a finite resource. Every minute spent answering "how many leaves do I have" is a minute not spent on the human, irreplaceable parts of the job - coaching managers, resolving real conflicts, designing better policies.

    • Employees deserve fast, correct answers. Not "let me check and revert." Not policy PDFs hidden on Keka.

    • HR deserves insights, not just questions. The dashboard turns every conversation into a signal. Which policy is most-asked? Which one is most misunderstood? Which grievance theme keeps coming up? HR now has data instead of anecdotes.

    In the first week of internal testing, employees asked Anuradha everything from "what's our POSH policy" to "I'm thinking about taking a sabbatical, what are my options" - and got grounded, accurate, on-brand answers in seconds.


    What I Learned

    A few things, in no particular order:

    1. The hardest part wasn't the AI - it was getting the avatar to not say "I can't share that PDF" while a download button rendered right below it. Prompt engineering is real engineering.

    2. Voice mode is magic but unforgiving. Web Speech API, Cartesia TTS, abortable streams, and "what if the user presses End while audio is mid-flight" - every one of those took longer than expected.

    3. Users will absolutely upload a PDF of their offer letter and ask a question about it within 30 seconds of discovering attachments. Build for that on day one.

    4. AI tools are only as good as the humans they free up. The point of Anuradha isn't to replace HR. It's to make sure that when you do talk to HR, it's about something that actually needs a human.


    Try It

    If you're curious, sign in with your work email and have a chat. Ask her about leaves, travel, POSH, or just say hi. She'll remember next time.

    And if you're Anuradha reading this - yes, she sounds exactly like you, and I'm sorry/you're welcome. 😄


    Built with ❤️, GPT-4o, and a probably-unhealthy number of late-night git pushes. Please reach out for testing it yourself :)

    ConveGenius Daily Signals

    Receive the next signal

    Get future product, design, AI, engineering, and team signals directly in your inbox. Only published signals. No spam.

    Unsubscribe anytime · No tracking pixels

    Reactions
    Sign in to react

    Discussion

    1 comment
    Login to comment. Use Google to join the discussion.
    Login to comment
    • Om Kumar
      Om Kumar5/15/2026

      A mind that never try to stop in terms of Innovation - Great blog 🙌🏻