← Back to library

Let's integrate AI Agents in Event-Sourced Systems — Divakar Kumar, FlyersSoft

Watch videoView transcript

AI Engineer21m 36sTranscript ✅Added Jul 30, 9:47 pm GMT+8

Analyzed: 2026-07-30

Actionable Insights

  1. Keep the agent as tier two. Retain rule/ML paths for clear approvals and declines; invoke the agent only between calibrated thresholds. Log the threshold, evidence retrieved, model/version, tool calls, verdict, confidence, and downstream decision. Evaluate false positives, false negatives, review latency, and cost specifically for gray-zone cases.
  2. Build read models instead of exposing raw event stores. Create purpose-built projections for transaction history, account/KYC, device trust, payments, and chargebacks. Feed them through narrow tools with typed schemas. This reduces cross-bounded-context coupling and prevents the agent from inventing joins or scanning unrestricted data.
  3. Use CDC with idempotency and freshness checks. Propagate approved fields from event streams to the semantic/read layer using change-data-capture or consumers. Include event IDs, schema versions, timestamps, and replay-safe upserts. Reject an agent decision when required projections are stale or incomplete.
  4. Bound the reasoning loop. Set maximum tool calls, time, spend, and retry count; define explicit stop states such as approve, decline, escalate, and insufficient_evidence. Require human review for high-value transactions and novel patterns. Test loops with missing tools, contradictory events, and timeouts.
  5. Replay historical cases before live traffic. Run the full orchestrator against a labeled event-history slice. Compare it with the incumbent rules/ML system and a no-agent baseline. Deploy in shadow mode, then a small canary; never let a demo bypass financial-model governance, privacy, or adverse-action requirements.

Core thesis

Do not replace deterministic fraud controls with an LLM. Route only ambiguous “gray-zone” cases to a bounded agent that reads projections from existing domain events, produces a reasoned recommendation, and leaves final policy enforcement auditable and reversible.

Big ideas / key insights

  • Keep the agent as tier two: Retain rule/ML paths for clear approvals and declines; invoke the agent only between calibrated thresholds. Log the threshold, evidence retrieved, model/version, tool calls, verdict, confidence, and downstream decision. Evaluate false positives, false negatives, review latency, and cost specifically for gray-zone cases.
  • Build read models instead of exposing raw event stores: Create purpose-built projections for transaction history, account/KYC, device trust, payments, and chargebacks. Feed them through narrow tools with typed schemas. This reduces cross-bounded-context coupling and prevents the agent from inventing joins or scanning unrestricted data.
  • Use CDC with idempotency and freshness checks: Propagate approved fields from event streams to the semantic/read layer using change-data-capture or consumers. Include event IDs, schema versions, timestamps, and replay-safe upserts. Reject an agent decision when required projections are stale or incomplete.
  • Bound the reasoning loop: Set maximum tool calls, time, spend, and retry count; define explicit stop states such as approve, decline, escalate, and insufficient_evidence. Require human review for high-value transactions and novel patterns. Test loops with missing tools, contradictory events, and timeouts.

The durable theme is agents in event-sourced fraud systems. The video is most useful as an architecture and risk discussion, not as proof that a product or pattern universally works.

Best timestamped moments with interpretation

  • 0:00 — [music] » Hello everyone. Thanks for joining. So, I think like finally we are at the last day of the conference. Personally, I had a great experience, learned a lot of new things. So, I believe by the end of the session you would Interpretation: this is a transcript-backed point; the operational implication is to convert it into a measurable control or experiment rather than accept the presentation claim at face value.
  • 0:00 — So what we are trying to do is like we we had built a system where we had both these tier one system which has this rule based or the traditional ML model and then we also had a tier two system which is agentic AI approach like mo Interpretation: this is a transcript-backed point; the operational implication is to convert it into a measurable control or experiment rather than accept the presentation claim at face value.
  • 0:00 — broker and those those events in turn will be processed by a worker process and then it will be reaching out to our projection layer. So, the idea is like we need to gather all the data from all these different contexts and to hav Interpretation: this is a transcript-backed point; the operational implication is to convert it into a measurable control or experiment rather than accept the presentation claim at face value.
  • 0:00 — tools and these are really specific to our business use case, so we are trying to move those into this tools. And we also have a be- behavior analysis tool. So So, there like we are analyzing the transaction patterns with two diff Interpretation: this is a transcript-backed point; the operational implication is to convert it into a measurable control or experiment rather than accept the presentation claim at face value.
  • 0:00 — [music] Interpretation: this is a transcript-backed point; the operational implication is to convert it into a measurable control or experiment rather than accept the presentation claim at face value.

Practical takeaways / recommended workflow

  1. Define the decision/task, risk level, owner, and measurable acceptance criteria.
  2. Capture the existing baseline for quality, latency, cost, failures, and human review time.
  3. Implement the smallest reversible version with typed inputs/outputs, logs, and least privilege.
  4. Replay representative historical cases plus adversarial and missing-data cases.
  5. Run shadow mode or a small canary; compare results and inspect every failure.
  6. Expand only when benefits persist without increasing downstream review, security, or operational burden.

Comment insights

  • 0 likes — @aiDotEngineer: This talk is from the AI in Finance track at World’s Fair 2026. In October, finance gets the whole conference: AI Engineer New York, Oct 12 to 14 at the Sheraton New York Times Square, is our finance focused event of the year. Banking, hedge funds, trading, insurance, accounting. In production use c

Comments are audience evidence, not factual verification. Where comments challenge missing demos, sales framing, tool efficiency, or implementation detail, that pushback lowers confidence in broad claims but does not by itself refute the underlying architecture.

Deep research on the creator’s main claims

The strongest claims were checked against named external sources below. Because the configured web-search provider was quota-blocked during this run, I used stable official specifications, project repositories, and standards already identifiable from the topic; I did not convert unverified snippets into facts.

Supporting evidence: these sources support the underlying patterns—explicit verification, event/read models, least privilege, provenance, graph retrieval, typed orchestration, distributed tracing, and measured rollouts.
Contradicting or limiting evidence: none of them establishes the speaker’s vendor-specific metrics or proves that the proposed architecture beats simpler baselines in every environment. Added complexity, stale data, extraction errors, policy bypasses, coordination overhead, and operational cost remain material counterweights.
Verified fact vs interpretation: transcript and frames verify what was presented; external documentation verifies that the referenced patterns/tools exist. Performance, prevalence, and superiority claims remain interpretations unless a reproducible benchmark or primary study is cited.

My verdicts on major claims

1. Agents can use event-sourced context to improve ambiguous fraud decisions

Verdict: Mixed-positive — medium confidence. Event history and domain projections can improve context, but an LLM recommendation is not inherently calibrated or causally reliable.

Practical takeaway: test this claim against a documented baseline and preserve rollback/approval paths.

2. Existing rule/ML systems should remain tier one

Verdict: Agree — high confidence. This is the safest and most economically sensible architecture presented.

Practical takeaway: test this claim against a documented baseline and preserve rollback/approval paths.

3. A semantic layer solves cross-context access

Verdict: Mostly agree — medium confidence. It simplifies retrieval, but creates freshness, lineage, access-control, and deletion obligations.

Practical takeaway: test this claim against a documented baseline and preserve rollback/approval paths.

4. Agentic fraud decisions can be productionized through orchestration

Verdict: Agree with strict conditions — medium confidence. Typed tools, audit trails, bounded loops, replay evaluation, and human escalation are prerequisites; the talk underplays regulatory and model-risk governance.

Practical takeaway: test this claim against a documented baseline and preserve rollback/approval paths.

Screen-level insights

  • 0:32 — youtube-extract/o6U_2vd967Y/frames/000_000032.jpg: work projects. So, what is it we are going to learn? So, we are going to learn how to integrate AI agents in your existing system. So, this system it could be an even so system or it could be an even driven system or it could be any architecture that your business has invested ov. The frame anchors the spoken claim to the visible slide/demo; it matters because reviewers can distinguish an architecture diagram, metric, policy screen, or live workflow from narration alone.
  • 3:36 — youtube-extract/o6U_2vd967Y/frames/006_000216.jpg: had this rule-based engine like uh 5 years before. And this rule-based engine was perfectly fine like it was working perfectly fine for few of the cases. And but but the problem with this rule-based engine is like the maintainability because the fraudsters are trying to get intru. The frame anchors the spoken claim to the visible slide/demo; it matters because reviewers can distinguish an architecture diagram, metric, policy screen, or live workflow from narration alone.
  • 4:36 — youtube-extract/o6U_2vd967Y/frames/008_000276.jpg: able to handle most of the transaction because it would fall below a certain threshold then we would approve the transaction and if it goes beyond a certain threshold we would be blocking those transaction. But majority of the transaction like few of the transaction like goes und. The frame anchors the spoken claim to the visible slide/demo; it matters because reviewers can distinguish an architecture diagram, metric, policy screen, or live workflow from narration alone.
  • 5:37 — youtube-extract/o6U_2vd967Y/frames/010_000337.jpg: handle few of the areas like that is the gray zone areas with the help of agentic AI processing. So this is the approach like that we decided okay let’s move on with this approach." But then our architecture, our domain is really complicated. So, this is our different bounded con. The frame anchors the spoken claim to the visible slide/demo; it matters because reviewers can distinguish an architecture diagram, metric, policy screen, or live workflow from narration alone.
  • 6:07 — youtube-extract/o6U_2vd967Y/frames/011_000367.jpg: transact. And everything related to a transaction would be residing on this particular context. And it doesn’t have any information about the customer it is handling, or it doesn’t have any information about the payments or the device details. That is what the bounded context mea. The frame anchors the spoken claim to the visible slide/demo; it matters because reviewers can distinguish an architecture diagram, metric, policy screen, or live workflow from narration alone.
  • 8:09 — youtube-extract/o6U_2vd967Y/frames/015_000489.jpg: And we also have an asynchronous way of communicating with within other different contexts like through a message broker. If I zoom in a little bit on the transaction context, you could see what what are the details that these transactions are holding. So, essentially like what h. The frame anchors the spoken claim to the visible slide/demo; it matters because reviewers can distinguish an architecture diagram, metric, policy screen, or live workflow from narration alone.

My read / why it matters

This talk matters because it shifts attention from model fluency to system design. The practical question is not whether an agent can produce a plausible answer once; it is whether a team can constrain inputs and authority, recover the evidence, detect failure quickly, and demonstrate better outcomes than a simpler workflow. Adopt the narrow mechanism that survives replay tests and operational review—not the conference-level slogan.

Verification notes

Four explicit passes were completed before publication:

  1. Source/evidence audit: major claims were separated into transcript-backed presentation claims, externally corroborated patterns, and unverified metrics. Named sources and direct links are included above.
  2. Transcript/comment/frame fidelity audit: timestamp selections come from extracted transcript chunks; comment summaries preserve their status as opinions; screen descriptions reference extracted frame paths and nearby transcript rather than guessing unseen UI details.
  3. Hallucination/overclaim audit: vendor statistics and universal superiority claims were downgraded where primary methods/results were unavailable. No install command or product capability was asserted without a stable source.
  4. Actionable Insights audit: all five top items specify a first implementation move, evaluation criteria, and a caution, prerequisite, or rollout boundary; generic summary bullets were rejected.

Residual uncertainty: automated captions may contain name/product errors; web search was quota-blocked; referenced tools were not executed in this analysis; and vendor-reported studies should be rechecked against their primary methodology before procurement or policy decisions.