After the Prototype - agents, applied AI, and models put to work

Preferred format: 30 minutes, talk without demo. Topic: Agents in production.

Abstract

We built LLM agents for wealth management: assistants that answer questions about portfolios, funds and market data for advisors and investors. Before launch they did well on our offline evals for retrieval quality, answer correctness and tone. Within weeks of real traffic we found failures our evals never covered: questions phrased in ways our test set didn’t include, tool calls that were technically valid but used stale or incomplete data, confident answers where the right move was to hand off to a human, and multi-turn conversations that drifted away from what the user had asked.

This talk walks through what going from prototype to production actually looked like in a regulated fintech setting:

• Where the agent failed in ways our evals didn’t catch, and how we found out (user feedback, traces, advisor escalations, spot audits).
• What we had to build around the agent that we hadn’t planned for: guardrails, fallbacks, escalation to a human, and monitoring that watches behaviour, not just latency.
• How we rebuilt the eval loop: turning production failures into regression cases, separating retrieval errors from reasoning errors, and deciding what a human reviews versus what an LLM judge scores.
• What it cost in latency, money and engineering time, and which trade-offs we would make differently.

Specific numbers and client details are anonymised where needed; the architecture, failure patterns and process are shared in full.

Why it matters

Most teams can get an agent to pass a demo or an offline benchmark. The hard part is the gap between that and real users. In finance that gap is a trust and compliance problem, not just a quality problem. The lessons apply to any team running agents on real traffic.

Who it is for

Engineers, ML/AI engineers and engineering leads who are shipping, or about to ship, LLM agents or RAG systems to real users. Assumes basic familiarity with RAG, tool calling and evals.

Key takeaways

1. A taxonomy of production agent failures that offline evals tend to miss, with real examples.
2. A practical loop for turning production incidents into regression evals.
3. The “around the agent” checklist: guardrails, escalation paths and behaviour monitoring we wish we had built first.
4. How to decide what humans review and what automated judges score, and where each failed us.

Outline (30 minutes)

1. (3 min) The setup: agents for wealth management and what “passing the evals” meant for us.
2. (7 min) First contact with real users: four failure patterns our evals missed.
3. (8 min) Building around the agent: guardrails, fallbacks, escalation to humans, and behaviour monitoring.
4. (7 min) Rebuilding evals from production: failure-to-regression loop, retrieval vs reasoning errors, human vs LLM-judge review.
5. (3 min) What it cost, and what we would do differently.
6. (2 min) Takeaways.
   Q&A as time allows.

Speaker

Kumar Ravi, CTO India, TIFIN (Mumbai). A former founder turned CTO with about 17 years in tech. He leads engineering for AI-native fintech platforms built on LLMs, RAG and agentic systems, focused on turning AI experiments into reliable production systems. He has built engineering teams from zero as CTO at several companies.
LinkedIn: https://www.linkedin.com/in/techsavyravi/
Prior talks: this would be his first public talk.

Disclosures

Speaker works at TIFIN; the talk is about engineering lessons, not a product pitch. Client and portfolio details are anonymised.