FreddyLoop: Closed-Loop AI & KB Optimizer

An autonomous closed-loop agent for Freshservice that turns failed AI suggestions and thumbs-down ratings into continuous knowledge base updates and measurable deflection gains.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for FreddyLoop: Closed-Loop AI & KB Optimizer

Inspiration

In modern IT Service Management (ITSM) platforms like Freshservice, AI deflection is marketed as the primary metric for operational efficiency. However, in practice, AI agents like Freddy operate without a closed-loop feedback mechanism:

  1. When an end-user downvotes an unhelpful answer, that failure signal is silently discarded rather than piped back into model optimization.
  2. Deflection rates hit a hard ceiling due to stale, conflicting, or missing Knowledge Base (KB) articles.
  3. IT teams lack granular visibility into which specific articles cause escalations.

We built FreddyLoop to bridge this structural gap, turning silent AI failures into autonomous, continuous knowledge base improvements.

What it Does

FreddyLoop acts as an autonomous supervisory meta-agent embedded directly within the Freshworks AI Agent Studio:

  • Failure Event Interception: Intercepts low CSAT ratings, "thumbs-down" interactions, and bot escalation triggers in real time.
  • Root-Cause Triage: Classifies failure modes into three categories: Stale Article, No Coverage, and Wrong Intent.
  • Autonomous Remediation: Gathers recent resolved ticket resolutions and auto-drafts corrected, 4-step verified KB articles.
  • Correction Brief & 1-Click Publishing: Surfaces an actionable brief for IT administrators to review, listen to via ElevenLabs TTS, and deploy to production with a single click.

Deflection Impact & Mathematical Formulation

To quantify the operational impact of continuous KB remediation, FreddyLoop models the target deflection efficiency \(D_{\text{eff}}\) as:

$$ D_{\text{eff}} = \frac{\sum_{i=1}^{N} (Q_i \cdot C_i \cdot \alpha_i)}{T_{\text{total}}} \times 100 $$

Where:

  • \(T_{\text{total}}\) represents total incoming employee service requests.
  • \(Q_i\) is the query volume matching topic \(i\).
  • \(C_i \in [0, 1]\) is the KB article accuracy index (updated continuously via automated triage).
  • \(\alpha_i\) represents the intent alignment coefficient.

By identifying and repairing broken articles (such as deprecated VPN setups), FreddyLoop achieved a +18 percentage point increase in deflection rate, taking prototype performance from \(43\%\) to \(61\%\).

How We Built It

  • Frontend Dashboard: Built with Next.js, React, and Tailwind CSS, strictly replicating Freshworks' proprietary Crayons design system and 3-column Agent Studio layout.
  • Autonomous Feedback Engine: Implemented workflow logic to categorize negative deflection logs and generate structured remediation drafts.
  • Voice Synthesis Integration: Connected the ElevenLabs API to deliver hands-free audio summaries of Correction Briefs for system administrators.
  • Deployment & CI/CD: Deployed natively on Vercel with version-controlled repository staging on GitHub.

Challenges We Ran Into

  • Designing Within Crayons Aesthetics: Accurately replicating the Freshworks enterprise UI system without access to private internal component libraries.
  • Triage Heuristics: Balancing sensitivity between genuine documentation drift and isolated user dissatisfaction without flooding administrators with false-positive briefs.
  • Zero-Latency Audio Delivery: Streaming synthesized voice buffers from ElevenLabs into the browser without blocking UI state transitions.

Accomplishments We're Proud Of

  • Designed a clean, native-feeling extension for the Freshworks Agent Studio in under 24 hours.
  • Solved two interrelated ITSM bottlenecks (Closed-Loop Feedback and Knowledge Base Decay) with a single unified supervisor architecture.
  • Delivered a fully deployed, interactive Vercel prototype featuring working root-cause diagnostic flows and audio briefing playback.

What We Learned

  • Feedback signals without automated remediation pipelines are lost operational value.
  • Autonomous agentic workflows must balance automation with administrative human-in-the-loop control (1-click approvals) to maintain enterprise trust.

What's Next for FreddyLoop

  • Stage 2 Production Hook: Connecting live Outbound Model Context Protocol (MCP) endpoints and webhooks directly to Freshservice production APIs during the 24-hour build in Bangalore.
  • Multi-Modal Ingestion: Automatically extracting step-by-step resolution screenshots from human agent chat sessions to attach to auto-generated KB drafts.
  • Automated Regression Audits: Proactively scheduling synthetic query probes against Freddy AI to detect outdated content before employees ever encounter it.

Built With

Share this project:

Updates

Submission history