All work

Built for production. Measured in outcomes.

OCR pipeline for Indian government documents

BFSIGovernment· Workflow automation

98%

accuracy

35%

faster

10+

doc types

Problem

Manual KYC document processing at scale

Financial institutions handling thousands of documents daily — manual extraction was slow, error-prone, and creating compliance backlogs.

Solution

Async GPU-accelerated OCR pipeline

PaddleOCR for multi-language extraction, FastAPI for routing, Celery workers for parallel processing, Redis-backed queuing, and Prometheus dashboards for real-time SLA monitoring.

Result

Production-deployed, fully observable

98% text extraction accuracy across all document types. 35% reduction in processing time. Horizontally scalable, deployed on AWS EC2 g4dn.4xlarge, handed over with full documentation.

StackPaddleOCRFastAPICeleryRedisAWS EC2 g4dnPrometheus
Read full case study

SupportIQ — Enterprise AI support agent

BFSIHealthcare· Custom AI agents

<200ms

cache hits

4-route

intent routing

DPDP

compliant

Problem

Tier-1 support at scale with compliance constraints

A support team handling thousands of queries monthly across INFO, TRANSACTIONAL, COMPLAINT, and OUT-OF-SCOPE intents — each requiring different escalation paths and compliance treatment.

Solution

LangGraph multi-agent system with hybrid RAG

SPLADE sparse retrieval combined with dense embeddings over Qdrant. Presidio PII redaction at input and output layers in subprocess isolation. 3-tier LLM fallback. Redis semantic caching at 0.85 cosine threshold.

Result

DPDP-compliant, audit-ready, production-deployed

Autonomous routing and resolution with full escalation context. Sub-200ms cache responses for repeated queries. Complete interaction audit trail. Compliant with DPDP Act by architecture, not policy.

StackLangGraphQdrantSPLADE RAGPresidioRedisFastAPIGroq
Read full case study

Mutual fund query chatbot

BFSI· Custom AI agents

Real-time

streaming

WS

protocol

Full

audit log

Problem

High query volume with financial compliance requirements

A mutual fund platform needed to handle investor queries about fund performance, NAV, and portfolio allocation — at scale, with full audit logging and financial guardrails.

Solution

WebSocket-based streaming chatbot

FastAPI WebSocket architecture for real-time response streaming. LLM with financial domain grounding. Structured response validation to prevent hallucinated figures. Complete conversation logging for compliance.

Result

Live, audit-compliant, zero-hallucination guardrails

Real-time streaming responses with sub-second time-to-first-token. Financial figures validated against live data before display. Full conversation audit trail meeting SEBI guidance on AI-assisted advice.

StackFastAPIWebSocketsOpenAIQdrantDBCalendly API
Read full case study