Varun Ungurala

• **90% ** — Reduction in Time from Idea to App
• **60% ** — Increase in User Task Completion Rate
• 40% — Reduction in Input Friction (Voice vs Text)

Users faced difficulties in transforming ideas into usable apps due to technical complexity. Existing tools required structured input, hindering accessibility for non-technical users, while multilingual users experienced friction in expressing intent clearly. Building a real-time AI generation system with reliability and usability posed additional challenges.

I designed a voice-first GenAI system that allows users to convert natural language into functional apps instantly. I developed an LLM orchestration layer to interpret user intent and generate structured outputs. Agentic workflows were implemented to break down user needs into executable steps. The system was equipped with multilingual support and a simplified UX for a zero learning curve. Feedback loops and iteration mechanisms were added to enhance output quality over time.

Reduced the time from idea to app, enabling near-instant app creation. Improved accessibility for non-technical and multilingual users, and set up a new interaction model for AI-driven app generation, leading to increased engagement and repeat usage.

• 50% — Faster AI Model Selection
• 30% — Reduction in AI Usage Cost
• 90% — Improvement in Output Evaluation Confidence

Teams faced difficulties in efficiently comparing outputs across multiple LLMs, leading to delays and inconsistent evaluations. There was a lack of visibility into cost, latency, and performance metrics, making it challenging to select the right model for specific use cases.

I built a multi-LLM comparison platform allowing for side-by-side output evaluation. The solution included API-based orchestration across OpenAI, Claude, Groq, and Mistral. I designed workflows to facilitate prompt testing, response comparison, and evaluation. Usage tracking was added for insights into cost, latency, and performance. The architecture was made extensible to support new models and features.

Reduced time spent on model evaluation and decision-making, improved confidence in selecting the right LLM for each task, and enabled cost optimization through better model visibility. Increased productivity for teams working with multiple AI models.

• 70% — Reduction in Manual Effort
• **92% ** — Accuracy of Retrieved Insight
• 50% — Faster Decision-making

• The client manually processed thousands of reports, tables, and charts, leading to delays and inconsistent insights.
• Extracting meaningful information from mixed-content documents caused high analyst effort and errors.
• Manual review slowed decision-making.
• Integrating AI agents was difficult due to security, access control, and scalability concerns.

• Analyzed workflows to identify high-effort tasks and define use cases like corporate intelligence and ESG tracking.
• Built a flexible GenAI platform with RAG pipelines, embeddings, and metadata enrichment.
• Created multi-agent React workflows for automated document analysis, reasoning, and structured output.
• Implemented LLM benchmarking, similarity scoring, and re-ranking for accuracy and reliability.
• Deployed solutions with observability, access control, and enterprise-grade security to enable scalable production-ready use.

Freed analysts from manual work, accelerated decision-making, improved AI accuracy and consistency, enabled scalable deployment without extra headcount, and ensured traceable, compliant outputs for reliable insights.

• 95% — Improved Precision in Complaint Categorization
• $8 Million — Yearly Savings in Operational and Settlement Costs
• 60% — Faster Complaint Review through AI Automation
• 40% — Increased Analyst Capacity via Automated Evidence Extraction

Manual complaint handling at financial institutions was slow, error-prone, and costly. There was regulatory risk, reputation liability, and a high operational burden due to undisclosed fees, service issues, and commission disputes. Additionally, case handlers were overburdened, which slowed resolution and reduced customer satisfaction.

I developed a multi-agent AI system to ingest complaint data, analyze cases in accordance with internal policies and external regulations, and automate investigations. It was integrated with existing systems via APIs, enabling scalable, compliant, and efficient case handling. This enabled auditors to trace decision-making, ensuring full audit ability, lineage, and explainability.

We enabled analysts to focus on high-value investigations instead of manual data gathering, improved audit readiness with clear reasoning, and reduced operational bottlenecks. We allowed the team to scale investigations without increasing headcount.

• 40% — Reduction in Analyst Research Time
• $2 Millon — Yearly Savings

The client tracked over 8,000 companies for ESG compliance, and analysts manually downloaded reports and read dense text, charts, and tables. The process was slow, error-prone, and required large analyst teams. There was a high risk of missing insights and delayed compliance reporting.

I developed an AI workflow to crawl, filter, and process thousands of reports automatically. I chunked documents, added metadata, and stored them in a vector DB. I also solved extraction from charts, tables, and image-heavy pages, adding an RAG agent with similarity scoring, re-ranking, and reference lineage.

I reduced analyst workload and significantly sped up ESG research, achieving 90% accuracy with metadata enrichment and providing traceable answers. I enabled scalable processing without adding headcount and provided audit-ready responses.

• 40% — Daily App Usage Boost
• 55% — Manual Tagging Time Reduction
• 30% — Increase in Conversions
• 45% — Return Users

Users struggled to decide what to wear and how to style existing clothes. They had no organized way to track their wardrobe, no personalized fashion recommendations, and no easy method to discover what items they were missing. Retail brands also lacked insight into user preferences, leading to low personalization and poor conversion rates.

I built an AI-powered mobile app that allows users to upload photos of their clothes, automatically classifying items using CNN models, and analyzing color, pattern, and style preferences. I designed a personalized recommendation engine that matched outfits, suggested new items based on wardrobe gaps, and generated a feed of influencers and celebrities for inspiration. I also added a digital closet, a QR code import system, and a virtual try-on experience for seamless engagement.

I enabled users to make better styling decisions, reduced wardrobe confusion, and improved discoverability. I provided valuable insights to brands about user preferences, unlocking more targeted marketing and higher conversion rates.