About the meet-up

Two years ago it was prompt engineering, then context engineering, then the harness. Then the loop. Now the graph — and state that has to survive a crash at step 40 of 60. The Fifth Elephant Hyderabad meet-up is a practitioner view of what is ceremony and what is real.

📅 Saturday, 5 September
🕑 10:00 AM – 12:45 PM
📍 Lloyds Technology Centre, Lobby 5 Floor 23, Sattva Knowledge Park — https://maps.app.goo.gl/3TzY5eHyXQKH7f5m9
🔴 Live stream available exclusively for Fifth Elephant members. Free in-person seats for everyone else.

The abstractions keep shifting — prompt, context, harness, loop, graph — and each wave arrives with a flood of posts from people selling a framework. It takes about a year to find out what actually held.

Meanwhile more teams are training and serving their own small models, and not all of them sit inside an agent. Sometimes it’s cost, sometimes it’s data that can’t leave the building.

Discussions at the meet-up will centre on some of these questions:

• Which agent abstractions are earning their keep, and which are ceremony?
• Is a tuned 8B good enough for your problem — how would you know?
• What does it take for state to survive a crash at step 40 of 60?
• What made inference cheap or fast enough to actually ship?

Featured talks

1. A green harness is not a healthy harness by Karun Japhet (Engineer & Consultant, Sahaj Software) on the guardrails you set once and never check again: standing-rules files that go stale until they’re confidently wrong, rule density that crosses an adherence cliff, and agents that can satisfy a failing gate by deleting the test that enforces it. Karun draws the line between guardrails the agent can quietly switch off and the ones that hold regardless — and gives three practices for treating the harness itself as production infrastructure.

2. Good enough for what? by Prudhvi Krovvidi (Data Scientist, Straive) on getting burned three times this year by the same mistake: checking the benchmark number everyone reports instead of the one the product actually needs. A local OCR model looked nearly as good as hosted frontier APIs on word accuracy — until he checked whether it could still read ID numbers, where it silently failed on most of them. Prudhvi will run the failure live on stage, and walk through two more cases where “good enough” turned out to be the wrong question.

3. Gantry: Openclaw for your company) by Ravi Kiran Vemula (VP of Engineering, KnackLabs) on building an agent that behaves like a coworker — one identity across Slack, Teams, and product channels, memory that follows the person, access granted upfront rather than negotiated per step. Ravi will walk through the harness pieces that held in production and the ones that didn’t: a permission classifier that blocked a tool call one hour and waved through the identical call the next, jobs that pause on denial and resume cleanly after a crash, and audit records that still make sense weeks later.

Editorial team

Curated by Chaitanya Sangani, Madhusudhan Sambojhu, Srix Sriramkumar, Prasun Anand, and guest reviewer Raghotham Sripadraj — practitioners across AI engineering, small language models, and The Fifth Elephant community.

💡 Why you should attend

• Real production lessons on agent harnesses, evals, and small-model deployment — not launch-post slides.
• Connect with practitioners across dev tooling, healthcare, legal, financial services, regional languages, and enterprise search.
• Walk away with a clearer read on which agent abstractions are worth adopting, and which aren’t yet.

🧭 How to attend

• In-person participation is free — just click Register to confirm your seat.
• Remote participation available for Fifth Elephant members. 👉 Get your membership from https://hasgeek.com/fifthelephant#memberships

🏢 Venue hosts

Thanks to Lloyds Technology Centre for hosting this edition in Hyderabad.

📞 Contact

📧 Email: info@hasgeek.com
📱 Call/WhatsApp: +91 7676332020