Case Studies

Five clients. Four countries. All solo.

Every engagement was owned end-to-end: discovery calls, technical scoping, architecture, implementation, and production deployment. No team. No handoffs.

🇮🇳🇦🇺🇮🇹🇬🇧

Products Shipped

Countries

AI Tutor Platform

🇮🇳India

Ed-tech client approached for consulting. Retained for full product delivery.

The Engagement

Client initially reached out for consulting advice only.
Conducted discovery sessions to understand the learning problem.
Scoped the MVP, estimated timeline, and effort independently.
Client trusted the vision and expanded the engagement to full delivery.
Ran weekly update calls throughout the build.
Launched MVP1 to real users and stayed on for the MVP2 roadmap.

What Was Built

Custom AI tutor agents with personalized feedback loops.
Session tracking and learning progress analytics.
Real-time AI responses with sub-second latency.
Admin dashboard for content and user management.
Authentication and role-based access.
Scalable backend architecture ready for MVP2 expansion.

Logistics Operations Platform

🇮🇳India

Enterprise-grade admin panel with real-time GPS, KYC workflows, and ONDC-compatible architecture. Delivered in 3 weeks.

Scope Delivered

Real-time dashboard with live operational metrics.
Order management with bulk assign and NDR handling.
Driver KYC workflow and scheduling system.
Live GPS tracking map with real-time driver locations.
Auto-assignment rules engine.
Financial management: payouts, commissions, and pricing configuration.
Analytics and reports with CSV export.
Role-based access control for Admin and Manager workflows.

Architecture Decisions

Designed an ONDC-compatible API structure so the client backend team could plug in later without rework.
Built a mock API server to decouple frontend progress from backend readiness.
Structured admin actions to support future ONDC update calls without schema changes.
Independently scoped the timeline, estimated effort, and delivered on schedule.

RAG Compliance Auditor

🇦🇺Australia

AI-powered compliance document analysis system for an Australian enterprise client.

The Problem

Client needed to extract and analyze information across large sets of compliance documents for audit reporting.
Manual review was slow, error-prone, and unscalable.
Compliance accuracy was non-negotiable and hallucinations were unacceptable.
The engagement had to run cleanly across Australian and Indian time zones.

What Was Built

RAG pipeline with Pinecone vector retrieval for semantic document search.
High retrieval accuracy on compliance-heavy documents.
Sub-200ms query latency in production.
Structured audit report generation from retrieved context.
FastAPI backend with rate limiting and production-grade error handling.

YouTube Automation Desktop App

🇮🇹Italy

End-to-end AI video creation tool for an Italian content creator. Script to publishable MP4, fully automated.

The Engagement

Client running a faceless YouTube channel needed the entire video production workflow automated.
Scoped the full tool architecture independently from a short client brief.
Delivered a native macOS desktop application instead of a web demo.
Client could go from script to publishable video without touching editing software.

What Was Built

AI voiceover using ElevenLabs TTS with voice customization.
Auto image generation mapped to script segments.
Video editing panel with subtitle overlay.
Background music and intro/outro insertion.
Direct render to MP4 for production-ready output.
Desktop GUI in Tkinter so the client never had to touch the command line.

AI Financial Document Automation

🇬🇧United Kingdom

AI-powered invoice processing and financial data automation integrated with Xero for a UK fintech client.

The Engagement

UK fintech client needed to automate manual invoice processing workflows.
Scoped the full solution independently: OCR pipeline, AI extraction, and Xero integration.
Served as the sole technical contact across communication and delivery.
Translated business requirements into technical architecture without a PM or BA layer.

What Was Built

OCR-based ingestion pipeline for invoice extraction.
AI-powered data extraction with prompt engineering, improving accuracy by 35%.
Full Xero API integration for automated financial data sync.
Robust API design with error handling and retry logic for production reliability.
End-to-end ownership from scoping and architecture to deployment.

StutterEase

🚀Startup

Co-founded and built a real AI product for people who stutter. 40+ active users, hospital-level customer discovery, and a conscious pivot.

The Journey

Identified a real problem: consistent and affordable speech therapy support is hard to access.
Visited speech therapy hospitals and clinics in person to observe the domain closely.
Ran structured customer discovery with therapists and patients before locking the product direction.
Built and launched to real users, reaching 40+ active users on the platform.
Iterated from user feedback and made a deliberate pivot based on business reality, not technical failure.

What Was Built

AI conversational platform with custom agents for real-time speech analysis.
Multi-step reasoning agents delivering personalized feedback after each session.
ASR model integration for live speech processing.
NLP pipeline for fluency pattern detection and structured coaching feedback.
Full production deployment on AWS with Docker containerization.
Working product demo used in real conversations with customers and partners.

AI Legal Composer

⚖️Startup

Co-founded and built an LLM-powered legal document generation system from scratch for a high-stakes professional domain.

The Problem

Legal drafting is slow, expensive, and inaccessible for many small businesses and individuals.
Generic LLMs hallucinate legal language and produce unreliable documents.
The system needed real domain adaptation to become trustworthy enough for use.
The entire product was built from idea to working system as co-founder and sole AI engineer.

What Was Built

LLM adaptation for the legal domain through prompt engineering and domain-specific fine-tuning.
Multi-component pipeline spanning transcription input, entity extraction, and structured document generation.
Transformer-based NLP pipeline converting client briefs into formatted legal outputs.
Accuracy-first system design where hallucination reduction was the core challenge.
End-to-end ownership across ideation, architecture, build, and iteration.