Voice response latency
~1-1.25s
Near real-time performance for natural phone interactions.
AI Workforce SaaS
Built a multi-channel AI workforce that answers, assists, and automates at scale.
AI Workforce SaaS is a production-grade platform that enables businesses to launch AI assistants across phone, SMS, and web chat, combining real-time conversational AI with workflow automation, business knowledge retrieval, and CRM-connected actions.
The core product promise was simple: help businesses automate front-desk, support, and lead
handling workflows 24/7 without sacrificing response quality or the speed users expect from live
conversations.
Availability
24/7
Always-on communication handling with zero waiting for staffed teams.
Channels unified
3
Phone, SMS, and web chat powered by a shared AI agent core.
Business outcome
10-50x ROI
Driven by automation, faster response times, and lower support load.
What it solved
- Missed calls and slow first-response times.
- Disconnected phone, SMS, and chat experiences.
- Low-quality generic bots with weak domain accuracy.
- Scaling roadblocks: Enabled massive scale for enterprises while giving SMBs a pay-as-you-go model with zero hiring required.
What was built
- Real-time voice AI pipeline with streaming responses.
- Reusable multi-channel agent engine.
- Knowledge ingestion, retrieval, and workflow automation.
Why it mattered
- Businesses could automate service without expanding headcount.
- Customers received fast, contextual, human-like assistance.
- The product scaled across multiple industries and use cases.
- Support availability: Made 24/7 coverage easy and affordable.
The Challenge
Businesses needed more than a chatbot. They needed an AI system that could handle real customer interactions across channels, keep latency low enough for voice, and execute business workflows reliably in production.
Operational constraints
- Real-time voice UX breaks down when latency feels conversationally slow.
- Multiple simultaneous calls require concurrency and fault tolerance.
- Phone, SMS, and chat each have different protocols and user expectations.
- Healthcare and service workflows required domain-sensitive behavior.
Product goals
- Deliver human-like AI conversations in voice and text.
- Automate front-desk, support, and booking flows end to end.
- Ground responses in business-specific knowledge and CRM context.
- Make assistant creation simple enough for business admins to self-serve.
The Solution
The platform was designed as a modular AI agent system: communication channels fed into a shared orchestration layer, which combined prompts, retrieval, workflow execution, and response generation to produce fast, grounded, action-oriented outcomes.
Core product flow
- Business admins create an AI agent and configure channels.
- Business URLs and uploaded files are scraped, parsed, and embedded into the knowledge base.
- The runtime engine interprets intent, retrieves context, executes actions, and responds.
- Integrations connect the assistant to CRM, booking, and internal business systems.
Design principles
- One agent core, many channels.
- Low-latency first for voice interactions.
- Retrieval-grounded responses to reduce hallucinations.
- Workflow completion over chatbot-style dead ends.
Capability Snapshot
The platform delivered more than AI conversations. It created a unified business automation layer that could answer questions, retrieve context, trigger workflows, and update external systems.
Conversation engine
Streaming STT, LLM token generation, and TTS orchestration for natural voice and text interactions.
Knowledge grounding
Website scraping, document parsing, chunking, embedding, and retrieval using a FAISS-backed vector store.
Workflow automation
Intent-driven actions for appointment booking, routing, ticketing, order handling, and CRM updates.
Multi-channel consistency
A shared context layer allowed the same business agent to operate across phone, SMS, and web chat.
Industry adaptation
Vertical-specific prompting and knowledge setup for healthcare, retail, finance, hospitality, and real estate.
Operational visibility
Conversation logs, latency tracking, escalation signals, and feedback loops for continuous quality improvement.
Implementation Highlights
Ultra-low latency voice pipeline Why it mattered: voice AI only works when it feels immediate.
Challenge: Traditional voice assistants often feel delayed and robotic, which makes conversations awkward and reduces trust.
Approach: The system streamed speech-to-text partials, generated incremental LLM responses, and coordinated text-to-speech output in parallel with optimized buffering.
Result: The platform achieved roughly 1-1.5 second response times, enabling natural, interruptible, human-like phone conversations.
Unified channel-agnostic agent core One intelligence layer for phone, SMS, and web chat.
Different channels create different UX constraints, but businesses need consistent behavior. A shared agent core was built beneath the communication layer so channel-specific inputs could map into the same memory, orchestration, and action systems.
This let the same assistant handle calls, messages, and chat sessions without rebuilding business logic for each channel.
Business knowledge ingestion and retrieval Grounded responses instead of generic chatbot answers.
The onboarding flow allowed businesses to submit websites, documents, FAQs, SOPs, menus, and other internal content. That data was scraped or parsed, chunked, embedded, and stored in a self-hosted vector layer for semantic retrieval at runtime.
Combined with CRM context from Zoho Desk, the assistant could answer with current, business-specific information and significantly reduce hallucination risk.
Workflow automation beyond conversation The assistant did work, not just Q&A.
The runtime followed an event-driven pattern: Trigger -> Intent -> Action -> Response. This enabled appointment booking, order processing, call routing, ticket creation, and CRM updates.
Integrations with scheduling systems, databases, and business APIs turned the assistant into an operational interface rather than a passive chat surface.
Industry-specific AI behavior Accuracy improved when the assistant understood the domain.
Generic prompts were not enough for production. The solution used vertical-specific prompting and business context for healthcare, retail, finance, hospitality, and real estate use cases.
This improved relevance, increased task completion, and produced more dependable outputs in real customer workflows.
Observability and continuous improvement Production AI needs measurement, not guesswork.
The platform logged conversations, failure states, escalation patterns, and latency metrics to support monitoring and iteration. Dashboards and feedback loops made it possible to refine prompts, workflows, and business logic over time as usage expanded.
Results & Business Impact
The strongest value came from combining speed, automation depth, and cross-channel consistency into one product that businesses could adapt to their own operational workflows.
Customer experience
- 24/7 responses across phone, SMS, and web chat.
- Faster lead handling and reduced drop-off from missed inquiries.
- More natural voice interactions through low-latency response design.
Operational efficiency
- Significant reduction in manual support and front-desk workload.
- Automation of booking, routing, and ticketing workflows.
- Ability to scale communication volume without proportional hiring.
Product differentiation
- Unified multi-channel AI instead of separate disconnected tools.
- Business-specific knowledge grounding and CRM-aware responses.
- High-value vertical fit across service-heavy industries.
My Role
This project involved end-to-end ownership across architecture, AI systems, backend engineering, integrations, and production reliability.
Owned directly
- System architecture for real-time AI and communication flows.
- LLM pipelines, prompting, orchestration, and optimization.
- Knowledge ingestion, retrieval, and vector search workflows.
- Multi-channel backend infrastructure for phone, SMS, and web.
- Workflow engine design and external integrations.
- Performance tuning for latency, concurrency, and reliability.
Hands-on execution
- Wrote production backend and AI pipeline code.
- Debugged live call flows, hallucinations, async issues, and race conditions.
- Implemented APIs and internal services used by the product.
- Supported deployment, monitoring, and incident response in production.
- Collaborated on admin UX, product strategy, and vertical-specific solutions.
Key Lessons
The project reinforced a practical principle: successful AI SaaS products win when they balance model quality with response speed, workflow completion, and operational reliability.
Performance is product
In voice systems, latency is not a backend metric. It directly shapes user trust and adoption.
Domain context beats generic intelligence
Vertical-specific prompts and retrieval delivered much stronger outcomes than a one-size-fits-all assistant.
Automation must complete the loop
The highest-value experiences came from triggering real actions, not simply returning well-worded answers.