PATHOS

Pathos is an AI-driven PR platform used by a firm serving thousands of customers. I joined when the system was a Streamlit proof of concept built by a previous developer — a working demo that generated PR angles from public data using LLMs, with no authentication layer, hardcoded prompts, and no architecture designed for production use. Within months, the tool began evolving into a client-facing platform serving real customers. I was responsible for the frontend throughout and took on identifying and resolving technical problems across the backend as they emerged. At various points I managed a junior developer, and I made architectural and engineering decisions under delivery pressure and a frequently incomplete specification. I was part of the team through the company's IPO.

The Newshook was the central artifact of the product — what the system produced and what clients consumed. In the PoC, it existed as a loosely structured dictionary. As the product scaled, different parts of the codebase developed divergent assumptions about what a Newshook contained, how it was identified, and what operations could be performed on it. Formalising the schema required auditing every place where Newshook assumptions were encoded — and then managing the transition across the frontend, backend services, and downstream integrations simultaneously in a live system. In applied AI systems, ontology stability is a first-class engineering concern. The entities your system produces need to be defined with the same rigour as your API contracts, and that rigour must be established before usage patterns cement implicit assumptions into structural debt.

As the platform matured, performance became the constraint. Early bulk processing required approximately 5 minutes per lead. I identified the bottlenecks, restructured the prompt workflows, and introduced concurrency. Processing time dropped to under 1 minute per lead. Large batches that previously required 2–3 hours could be processed in 15–20 minutes. This required separating enrichment tasks from generation tasks, managing OpenAI token usage at scale, handling rate limits predictably, and translating raw transcripts into LLM-efficient representations. I also built search indexing via Algolia across approximately 200,000 leads and contributed to pipeline infrastructure for bulk ingestion.

In parallel with the core platform, I architected Pressella from scratch — a meeting ingestion agent built on recall.ai. Pressella joined sales calls, captured transcripts, enriched them with external context, and generated customised follow-up articles and emails. I designed the pipeline as a sequence of distinct LLM calls rather than a single monolithic prompt: transcript preprocessing, context enrichment, article angle generation, article drafting, and email synthesis with conditional emphasis. A post-call feedback layer was introduced, allowing salespeople to provide structured input after the meeting that shaped the generated output. Reliability was the primary design constraint — the system was architected from scratch with that priority, not retrofitted.

Not everything worked. Pressella failed Zoom Marketplace review due to a production configuration mismatch — the system functioned correctly in development but did not meet the deployment environment's compliance requirements. I adjusted the configuration flows, security posture, and production validation process. On bulk processing infrastructure, some solutions were pragmatic rather than elegant — distributed execution under time pressure. Governance and alerting matured more slowly than core functionality. That was a deliberate tradeoff given team size and delivery constraints, though earlier investment in monitoring would have reduced operational friction. Scope discipline was also a constant: protecting architectural boundaries required saying no to ideas that added novelty at the cost of stability.

• Platform cited as a proprietary technology asset in Pathos' AIM admission documentation.
• Per-lead processing time reduced from ~5 minutes to under 1 minute; large batch jobs from 2–3 hours to 15–20 minutes.
• Pressella operating as a structured ingestion layer within the full ecosystem.
• Auth, user management, and backend services stable under production load.
• ~200,000 leads indexed in Algolia across the platform's CRM-like workflows.