Why Your AI Support Bot Hallucinates — And How We Built an Evidence-Gated Fix

Ketan Chokkara — Sat, 07 Nov 2026 11:31:40 +0000

Building RelayDesk: Evidence-Gated Multi-Agent Support Triage on Switch AI

By Ketan Kumar Chokkara — AI Engineer & Data Scientist from India (@CodeUltr0n) — Built for the Flip the Switch Hackathon (Nov 4–7, 2026)

📺 Watch the 4-Minute Demo Video on YouTube: https://youtu.be/Gi6qcB_jUVQ

1. The Frontline Support Dilemma

Every software engineering organization faces a compounding crisis at the intersection of customer support and core engineering:

Over-escalation & Alert Fatigue: Support teams, overwhelmed by incoming tickets and lacking engineering context, reflexively escalate routine issues, transient glitches, or already-reported bugs to on-call engineers. Developers spend up to 30% of their sprints sifting through duplicate tickets.
Under-escalation & Hallucinatory Support: Organizations that deploy off-the-shelf generative AI chatbots often discover an even worse problem: ungrounded LLMs invent plausible-sounding troubleshooting steps, promise features that don't exist, or fail to recognize severe novel regressions that urgently require engineering intervention.

When building RelayDesk for the Flip the Switch Hackathon, our guiding principle was simple:

Never let an AI make a customer-facing or engineering decision without citing verified, structured evidence, and never take action without a human in the loop.

Here is the story of how we built RelayDesk — an evidence-gated customer support triage and engineering escalation system powered by Switch AI, Antigravity, and Discord.

2. System Architecture: The Four Outcome Tracks

At the core of RelayDesk is a deterministic, evidence-gated pipeline that classifies every incoming customer report into one of exactly four outcomes:

flowchart TD
    A["Support Lead / Customer Report"] -->|"Free Text"| B["normalize.py (Structured Extraction)"]
    B -->|"IssueReport"| C["retrieval.py (Deterministic Index)"]
    D[("Knowledge Corpus: Tickets, Runbooks, Releases, Logs")] --> C
    C -->|"Top-3 Candidates"| E["scoring.py (Feature Blending)"]
    E -->|"Scored Candidates"| F["decision.py (Threshold Routing)"]
    F -->|"Outcome: ANSWER/CLARIFY/DUPLICATE/ESCALATE"| G["handoff.py (Message Card Rendering)"]
    G -->|"Switch Room Feed"| H{"Human-in-the-Loop Approval"}
    H -->|"Approved Fix"| I["Customer Message Dispatched"]
    H -->|"Approved Duplicate"| J["Linked to Active Ticket"]
    H -->|"Approved Escalation"| K["sandbox.py (Isolated Repro Execution)"]
    K -->|"Reproduction Evidence"| L["Engineering Ticket Packaged"]

The 4 Decisions:

ANSWER (Confidence $\ge 0.65$): The issue is a known operational question with a verified resolution in a published troubleshooting runbook. The system cites the exact runbook section.
CLARIFY (Ambiguous / Missing Signals): Key diagnostics (component name, reproduction steps, version, or exact error text) are missing. The system asks a single targeted clarifying question rather than guessing.
DUPLICATE (Confidence $\ge 0.75$): The issue matches an active, open incident ticket (INC-xxxx) based on component, version, and error signature. Links to the existing incident to prevent duplicate triage.
ESCALATE (Novel Defect): The issue represents a genuine novel regression or unhandled error not covered by existing tickets or runbooks. Hands off to the investigation agent for sandboxed test reproduction.

3. Why Switch AI is Essential

RelayDesk is not a standalone Python script or a solitary bot; it is a collaborative multi-agent workflow designed for human teams. Switch AI provides the crucial infrastructure that makes this work:

Multi-Agent Specialization

Rather than asking one monolithic LLM to do everything, RelayDesk connects two specialized agents in a shared room bridged to Discord:

@relaydesk-triager (Frontline Triage): Configured with public addressing (anyone). Listens to customer and support lead messages in #support-triage, runs regex normalization, queries the deterministic retrieval index, scores evidence, and formats structured [EVIDENCE] and [DECISION] cards.
@relaydesk-investigator (Deep Investigation): Configured with restricted addressing (ownerAndAgents). Wakes up only when tagged by human operators or delegated by @relaydesk-triager. Its sole job is to spin up an isolated sandbox, execute reproduction test suites, and report verified stack traces.

Human-in-the-Loop Safety by Design

Every triage recommendation posted to the Switch room ends with:

Next: human approval required before anything is sent or merged.

No email is sent, no customer reply is dispatched, and no Jira ticket is filed until a human support lead gives the green light. Switch turns Discord into a mission control room where humans supervise AI agents in real time.

4. Under the Hood: Deterministic Scoring & Sandboxed Reproduction

1. Weighted Evidence Scoring Engine

Instead of asking an LLM "Does this ticket look like a duplicate?", RelayDesk computes an explainable 4-feature weighted score ($S \in [0.0, 1.0]$):

$$S = 0.35 \cdot M_{\text{error}} + 0.25 \cdot M_{\text{comp_ver}} + 0.20 \cdot M_{\text{timing}} + 0.20 \cdot E_{\text{repro}}$$

Error Symptom Match ($M_{\text{error}}$): Token containment between the normalized error signature and document tokens.
Component & Version Match ($M_{\text{comp_ver}}$): Strict metadata alignment with fallback to document body scanning.
Release Timing Match ($M_{\text{timing}}$): Proximity between the incident observation and service deployment dates.
Reproduction Evidence ($E_{\text{repro}}$): Quality of verified reproduction data (differentiating active open tickets from closed historical ones).

2. Disposable Sandbox Execution

When a novel defect is escalated, @relaydesk-investigator never runs tests in production or inside the main codebase. Instead, relaydesk/sandbox.py:

Copies the suspect service into a disposable temporary directory (tempfile.mkdtemp()).
Restricts commands strictly to allow-listed binaries (pytest, python, python3, make) parsed via shlex.split, running with shell=False to eliminate shell injection vulnerabilities.
Enforces strict execution timeouts (timeout_s: 120) and proxy blackholing for network isolation.
Caches and truncates test output tails (max_log_echo_lines: 20) to prevent log flooding and prompt injection.

5. Live Discord Walkthrough

📺 Watch the live multi-agent execution in action: https://youtu.be/Gi6qcB_jUVQ

Here is what the workflow looks like in practice inside our bridged Discord channel:

Scenario 1: Automated Duplicate Triage

Human:

@relaydesk-triager Customer reports payment total mismatch (HTTP 500) on checkout-service v2.14.0 since this morning's release.

@relaydesk-triager:

Parsed report:
Component: checkout-service
Version: v2.14.0
Error signature: payment total mismatch (HTTP 500)
Observation timing: Since this morning (following this week's release)

[EVIDENCE] Candidate shortlist for this report:
Checkout HTTP 500 — payment total mismatch after v2.14.0 rollout [ticket] score=0.86 id=INC-1042
why: Score 0.86 for [INC-1042] (ticket): driven by error_symptom_match=1.00, component_version_match=1.00.

[DECISION] DUPLICATE (confidence 0.86)
Why: Top candidate [INC-1042] is an active ticket with score 0.86 >= duplicate threshold 0.75.
Evidence: INC-1042
Next: human approval required before anything is sent or merged.

Scenario 2: Novel Bug Escalation & Sandboxed Reproduction

Human:

@relaydesk-triager Customer reports discount calculation failure in price_calculator on 10% discount orders. checkout-service v2.14.0. No matching past incident. Please escalate to @relaydesk-investigator to run reproduction tests in the sandbox.

@relaydesk-triager:

Evaluates candidates (top ticket score 0.70 < threshold 0.75). Issues [DECISION] ESCALATE and tags @relaydesk-investigator.

@relaydesk-investigator:

[REPRODUCTION RESULT] Bug Confirmed

Component: checkout-service (v2.14.0) — price_calculator.py
Reproduction: Reproduced in isolated sandbox (disposable copy, network-isolated, allow-listed pytest).
Root Cause: Discount quantizing uses ROUND_DOWN instead of ROUND_HALF_UP, causing a 1-cent calculation error on 10% discount orders (134.98 vs expected 134.97).

Test failure log tail:
>       assert result == Decimal("134.97"), f"Got {result}, expected 134.97"
E       AssertionError: Got 134.98, expected 134.97

Recommendation: Create an engineering ticket for checkout-service to switch decimal rounding to ROUND_HALF_UP in price_calculator.py. Awaiting human operator approval.

Scenario 3: Grounded Runbook Resolution (ANSWER)

Human:

@relaydesk-triager Customer reports stale search results in search autocomplete on search-service v3.2.1. Products that were deleted yesterday still appear in search results.

@relaydesk-triager:

Matches runbook search_troubleshooting (RB-501) and resolved ticket INC-1038. Outputs [DECISION] ANSWER (grounded in runbook RB-501). Drafts verified operational steps: manual Elasticsearch alias swap (POST /_aliases) and autocomplete cache TTL invalidation, pausing for human approval.

Scenario 4: Smart Diagnostic Clarification (CLARIFY)

Human:

@relaydesk-triager Customer says the CSV export button in Reports does nothing when clicked. No error message shown. No app version or browser provided.

@relaydesk-triager:

Refuses to hallucinate speculative answers. Detects missing technical context and outputs [DECISION] CLARIFY with a single targeted diagnostic question requesting app version, browser details, and developer console logs.

6. Rigorous Benchmark Evaluation

To prove that RelayDesk performs reliably beyond one-off demos, we built an automated benchmark suite (scripts/run_eval.py) evaluated across 8 synthetic incident cases split into development and held-out test sets, backed by 27 passing unit and integration tests:

Metric	Null Baseline (Always Escalate)	RelayDesk (Dev Split)	RelayDesk (Held-Out Split)
`correct_next_action`	25.0%	100.0%	100.0%
`escalation_precision`	25.0%	100.0%	100.0%
`escalation_recall`	100.0%	100.0%	100.0%
`top3_evidence_recall`	50.0%	100.0%	100.0%
`unsupported_assertion_count`	0	0	0
`median_latency_ms`	0.0 ms	0.5 ms	0.4 ms

Key Findings:

100% Escalation Precision: No non-bugs or known duplicates were sent to engineering.
100% Escalation Recall: Every single novel defect was caught and reproduced.
Sub-Millisecond Overhead: The scoring and routing logic takes under 0.5 milliseconds, adding zero perceived latency to the chat interaction.

7. Key Takeaways & What's Next

Determinism + AI Beats Pure LLMs: Using fast, deterministic heuristics for evidence retrieval and scoring prevents hallucinated citations and gives human operators full explainability.
Switch AI Makes Multi-Agent Real: Managing agent identities, addressing scopes, and Discord bridges via Switch turned a complex multi-agent architecture into a seamless team room experience.
Safety First: Sandboxed reproduction ensures agents can run automated diagnostic tests without risking the host system or production data.

Try it out!

Demo Video (YouTube): https://youtu.be/Gi6qcB_jUVQ
GitHub Repository: https://github.com/CodeUltr0n/RelayDesk
Quickstart: Run make test and make eval-heldout in under 30 seconds!

Thank you to WeMakeDevs and Switch AI for hosting the Flip the Switch Hackathon!

DEV Community: Ketan Chokkara

Why Your AI Support Bot Hallucinates — And How We Built an Evidence-Gated Fix

Building RelayDesk: Evidence-Gated Multi-Agent Support Triage on Switch AI

1. The Frontline Support Dilemma

2. System Architecture: The Four Outcome Tracks

The 4 Decisions:

3. Why Switch AI is Essential

Multi-Agent Specialization

Human-in-the-Loop Safety by Design

4. Under the Hood: Deterministic Scoring & Sandboxed Reproduction

1. Weighted Evidence Scoring Engine

2. Disposable Sandbox Execution

5. Live Discord Walkthrough

Scenario 1: Automated Duplicate Triage

Scenario 2: Novel Bug Escalation & Sandboxed Reproduction

Scenario 3: Grounded Runbook Resolution (ANSWER)

Scenario 4: Smart Diagnostic Clarification (CLARIFY)

6. Rigorous Benchmark Evaluation

Key Findings:

7. Key Takeaways & What's Next

Try it out!