About Customer:
Vahak is India’s leading digital transport marketplace, connecting shippers, transporters, brokers and truck owners on a single platform without intermediary commissions. Through its web and mobile applications, users post loads, attach vehicles, discover verified partners and track shipments in real time across a pan-India network.
The platform is built for a market where road carries more than 70% of goods movement and where logistics costs run at 13–15% of GDP against a global benchmark of 8–10%. Vahak’s focus is on bringing small and medium enterprises — the operators who have historically been locked out of large freight networks — into nationwide logistics transactions with transparent pricing and direct access to counterparties.
Industry:
Industries, Others
Offering:
Offerings, Gen AI/AIML
Business Challenges:
- As Vahak scaled nationally, support queries on load matching, payment status, vehicle operations and platform navigation grew faster than the support team could absorb. Around 80% of the incoming volume was repetitive and resolvable without a human agent, yet every ticket still landed in the same queue – pushing up response times across all query types.
- Response times lengthened as agents worked a single serial queue. Transporters waiting on a load match or a payment confirmation had no way to get an answer outside working hours, and slow resolution is a direct driver of platform abandonment in a marketplace where users can switch to a competing broker in minutes.
- Support was available in English only, while more than 60% of India’s truck operators and SME transporters are Hindi-first users. Large segments of Vahak’s own user base could not self-serve at all, so queries that should have been answered in seconds became agent escalations – and literacy barriers made text-only support unusable for a section of truck owners entirely.
- Answers varied by agent. Without a single grounded source of truth over Vahak’s FAQs and SOPs, responses to the same question differed between agents and shifts, and any generative approach risked inventing policy detail on payments and load handling that Vahak would then have to retract.
- Payment and load status checks required manual lookups. Agents moved between systems to answer “where is my payment” and “what happened to my load”, and unresolved issues depended on an agent remembering to raise a ticket – producing missed escalations, delayed payment resolutions and SLA breaches with enterprise clients.
- Support cost scaled linearly with the user base. At roughly $6 per human-handled interaction against $0.50 for AI-assisted resolution, every increment of platform growth added headcount, and peak-season freight surges had no elastic capacity behind them.
Solution Implemented:
- Rapyder architected and deployed a dual-runtime conversational support platform on AWS: a text runtime using an AgentCore-compatible invocation API with Strands-based orchestration for tool selection, response generation and escalation decisions, and a separate voice runtime for low-latency conversational audio – both operating on one unified decision layer.
- Built a real-time inference path tuned for conversational latency: Amazon Nova Sonic 2 bidirectional WebSocket streaming on the voice side, parallel guardrail and inference execution on the text side, retrieval pre-filtering to cut ranked candidates, and DynamoDB DAX for session reads.
- Delivered English, Hindi and Hinglish across both text and voice channels, so a truck owner can speak a query in Hindi and receive a spoken answer – removing both the language and the literacy barrier in a single interaction path.
- Grounded every answer in a Bedrock Knowledge Base built over Vahak’s own FAQs and SOPs, using hybrid retrieval with filtering and ranking, with Amazon Bedrock Guardrails applied at both input and output for PII redaction, topic blocklisting, hallucination detection and prompt-injection defence.
- Implemented deterministic tool calls for transactional data – payment and load status read directly from Amazon OpenSearch rather than inferred by the model – plus automated escalation that detects unresolved queries and payment issues and creates structured tickets through Vahak’s backend APIs, with retry, exponential backoff and a dead-letter queue for failed payloads.
- Deployed on serverless, event-driven infrastructure with per-session microVM isolation and auto-scaling runtimes, DynamoDB session state with TTL for cross-turn continuity, multi-AZ deployment, and OpenTelemetry tracing into CloudWatch GenAI dashboards with AgentCore Evaluations scoring live traces for correctness, helpfulness and goal success.
Services Used:
- Amazon Bedrock
- Amazon Bedrock AgentCore
- Amazon Bedrock Knowledge Bases
- Amazon Bedrock Guardrails
- Amazon Nova Sonic 2
- Amazon OpenSearch Service
- Amazon DynamoDB
- AWS Lambda
- Amazon API Gateway
- Amazon S3
- Amazon Cognito
- Amazon CloudWatch
- AWS WAF
- AWS Secrets Manager
- AWS IAM
- Rapyder GenAI Advisory & Solution Design
- Rapyder Managed Services
Business Benefits:
- 50% – Support Workload Reduced – Routine query automation
- 60% – Faster Response Times – Real-time AI inference
- 55% – Better User Experience – Multilingual voice + text
- 45% – Higher Answer Accuracy – RAG-grounded responses
- 70% – Concurrent Demand Scale – Cloud-native architecture
Ready to transform your customer support with AI?
Talk to Rapyder's AI team to see how an AI support platform on Amazon Bedrock can cut your workload and your cost per resolution.
Frequently Asked Questions
- Vahak deployed a conversational AI support platform built by Rapyder on Amazon Bedrock, resolving routine load, payment and platform queries autonomously across text and voice. Automation of repetitive queries cut support workload by an estimated 40–50%.
- The platform runs on Amazon Bedrock for foundation model inference and RAG, Bedrock AgentCore for agent runtime and orchestration, Amazon Nova Sonic 2 for voice streaming, OpenSearch for transactional lookups, DynamoDB for session state, and Lambda, API Gateway, S3 and CloudWatch across the supporting stack.
- Yes. The platform handles English, Hindi and Hinglish across both text and voice channels. Voice-first interaction is central to adoption among truck owners and small operators for whom text-only, English-only support was inaccessible.
- Retrieval-Augmented Generation grounds every response in Vahak's own FAQs and SOPs rather than the model's parametric knowledge. Hybrid retrieval with filtering and ranking, combined with Bedrock Guardrails and hallucination detection, lifted answer accuracy by an estimated 35–45%.
- Transactional data is retrieved deterministically from Amazon OpenSearch and formatted into a conversational reply. The model never infers a payment or load status, which removes an entire class of hallucination risk on the queries that matter most commercially.
- Each user session runs in a dedicated microVM that is wiped on session end, preventing cross-user data leakage. TLS 1.2+ in transit, AES-256 at rest, VPC-restricted OpenSearch, WAF and throttling at API Gateway, Secrets Manager for all credentials, and full OTEL audit trails in CloudWatch complete the control set.
- Yes. The agent-and-tool architecture is domain-agnostic: swapping the knowledge base and tool contracts adapts it to any high-volume support operation with repetitive queries, multilingual users and transactional lookups behind them.
