Most repetitive queries are now self-served through the bot, reducing dependency on manual HR responses across departments.
Kodertal delivers production-ready AI solutions, from strategy to scale. We design, build, and integrate AI agents, LLMs, chatbots, and automation that fit your workflow, ship faster, and drive measurable business outcomes globally.

Backed by vector search, SSO, and sentiment logic—no more missed queries or manual lookups.

Our client’s HR team faced rising volumes of internal queries on policies, benefits, and compliance—all requiring manual lookup and response. Employees lacked a direct way to get reliable answers without digging through PDFs or intranet links. We developed a custom HR chatbot powered by GPT-4o and Elasticsearch with vectorized semantic search. It understands employee intent, fetches relevant clauses, and auto-deflects repetitive queries. For complex cases, it triggers a live-agent handoff and logs the full transcript for review. Role-based access ensures responses are scoped and secure.
Industry:
Construction
Region:
California, USA
Company Size:
500+ Employees
Engagement:
12 Months
Fine-tuning GPT-4o embeddings to differentiate vague HR intents like “leave rules” vs “leave application” required multiple RAG iterations.
Applying correct response limits via SSO and RBAC without breaking query accuracy was a persistent cross-auth issue during development.
Semantic matches worked well broadly, but clause-level retrieval accuracy dropped without ELSER-driven field-aware filtering and query parsing logic.
Bot sometimes escalated chats too early or too late due to poor sentiment scoring calibration and missing fallback intent triggers.
Webhooks occasionally failed to sync ticket IDs or fetch updates due to timing mismatches in concurrent backend service execution.

To resolve ambiguity between overlapping HR queries, we trained GPT-4o embeddings on intent clusters and layered fallback prompts to improve routing accuracy across similar question patterns.
We built a dynamic prompt layer that injects user role and permissions into GPT inputs, ensuring answers remain scoped to the user's HR access profile without breaking intent alignment.
By integrating Elastic’s ELSER, we improved clause-level accuracy through sparse vector encoding, enabling the chatbot to retrieve specific content from complex HR policies and documents.
To avoid premature or delayed human escalation, we merged confidence scores and sentiment indicators, triggering escalation only when both thresholds signaled uncertainty or user dissatisfaction.
We built retry queues for ServiceNow and Ivanti webhooks, ensuring ticket actions are retried on failure and synced reliably—even under variable API response conditions.
300+
Lorem ipsum dolor sit amet consectetur.
6+
Lorem ipsum dolor sit amet consectetur.
50+
Lorem ipsum dolor sit amet consectetur.
20+
Lorem ipsum dolor sit amet consectetur.
GPT-4o-powered intent detection
Vector-based semantic HR search
Elastic ELSER document parsing
SSO-based role-level responses
Live agent switch via Connect
Real-time sentiment monitoring
Auto-ticket creation in ServiceNow
Dashboard with KPI visualization
Low-latency enterprise integration
React
Elasticsearch
Nest Js
Tailwind CSS
Open Ai
PostgreSQL
Python
Azure
HighLevel
Most repetitive queries are now self-served through the bot, reducing dependency on manual HR responses across departments.
The system resolved most HR-level queries end-to-end without escalation, deflecting tickets and easing helpdesk volume.
Combining GPT-4o and ELSER returned specific, clause-level answers to employee questions with minimal false positives.
Elastic indexing and vector matching returned answers within seconds—even under simultaneous multi-user load.
Role-based SSO filtering enforced scoped outputs, preventing access to out-of-bounds or sensitive HR content.
Chat transcripts and dashboard showed a clear shift in satisfaction, with most users rating interactions as smooth and helpful.
50%
Faster Query Handling
40%
Reduced HR Load
95%
Intent-Aware Responses
60%
Self-Served Queries
Projects Delivered
Full-Time Engineers
Years of Consistent Delivery
Technologies Mastered
Specialized team with 10+ years in machine learning and 5+ years in LLM development.
Specialized team with 10+ years in machine learning and 5+ years in LLM development.
Specialized team with 10+ years in machine learning and 5+ years in LLM development.
Specialized team with 10+ years in machine learning and 5+ years in LLM development.
