Case Studies

Context

TurboLegi supports staff who need to query, summarize, and cross-reference large volumes of legislative documents quickly. Users required AI directly in their existing workflow without switching tools. The challenge was integrating a reliable LLM layer in a professional environment with strict accuracy and attribution requirements.

Constraints

• Eight-week timeline from kickoff to deployment
• Integration had to work with existing authentication and identity systems
• Non-technical users required a zero-training interface
• LLM outputs had to be clearly distinguished from source documents

Approach

The system was broken into three components: the data pipeline, the interface, and the integration layer. Each was scoped independently with explicit integration points. The interface used a single interaction pattern to reduce cognitive load, and outputs were formatted to make provenance clear.

Technology

Backend used retrieval-augmented generation. Document embeddings were stored in a vector database, and LLM calls were contextually constructed from retrieved chunks. The API was stateless for scalable deployment. Frontend components followed existing conventions to minimize user disruption. Token usage was optimized by using docx mailmerge for document generation, rather than feeding entire documents into the ChatGPT API.

Outcome

Deployed on schedule with immediate adoption and no formal training. Two workflow gaps were addressed in a follow-up iteration. The retrieval pipeline reduced hallucinated citations compared to direct prompts, meeting the client’s key success metric.

Context

Gently needed to reach a large, fragmented consumer audience with targeted offers at scale. Existing tooling could not support the required personalization or send volume without significant manual effort. The solution needed to feel individualized while operating as a fully automated system.

Constraints

• High-volume outreach within a fixed launch window
• Dynamic personalization driven by user interest and activity signals
• Strict deliverability and sender reputation requirements
• Full handoff to the client team after launch

Approach

I reverse engineered a popular clothes resale website’s APIs to extract the necessary activity and interest signals. These signals were fed into a Rust-based pipeline that handled segmentation, variant selection, send pacing, and retries. Messaging was sequenced rather than broadcast to control throughput and reduce spam risk.

Technology

The system was built as an efficient, async Rust application optimized for concurrency and low memory overhead. Custom API clients interfaced with internal endpoints discovered through reverse engineering. The pipeline handled enrichment, template rendering, scheduling, delivery tracking, and real-time metrics in a single flow.

Outcome

The campaign generated over 5 million impressions. Personalized variants consistently outperformed generic sends across open and conversion metrics. After a short handoff period, Gently’s internal team operated the system independently and reused it for future campaigns without additional engineering support.

Context

Existing AI tools for students either encouraged over-reliance or provided no guardrails. AskAI needed to support multiple academic use cases while keeping students in control of output.

Constraints

• Multiple use cases in a single coherent interface
• Performance across a global user base with variable connections
• Academic, trustworthy experience rather than generic chatbot
• Simultaneous design and development with no hand-off phase

Approach

Each use case was implemented as a distinct interaction mode. Users could move between modes without leaving the session. AI prompting ran server-side to maintain consistency and prevent injection. Outputs were shown alongside original input for easy comparison.

Technology

Full-stack web app with server-side rendering for fast initial loads. AI features used a structured prompt layer and response streaming. Authentication, session management, and usage tracking supported multi-user scaling.

Outcome

AskAI launched globally within months. Users appreciated seamless switching between drafting and grading modes. Responsible-use design differentiated the product and eased institutional adoption.