Krishna – AI agent development, Typescript, React

Built a 0-to-1 AI product that interviews people across an organisation and converts conversational input into structured models of the business: organisations, people, processes, systems, data and information flows. The resulting operating model can be explored visually and used to identify friction, understand how work actually happened, and support AI-assisted workflow improvement.

• Designed and built multi-agent interview workflows that dynamically explored a business rather than following a fixed questionnaire, allowing the system to progressively discover people, processes, systems, dependencies and missing information.
• Built voice/avatar-based conversational experiences for conducting reverse interviews, combining real-time conversational UX with structured extraction so useful business knowledge was captured without asking users to fill in complex forms.
• Developed structured extraction and modelling pipelines that transformed unstructured interview conversations into organisation, process, data and flow models while preserving relationships and source context.
• Built interactive visualisations and a consultant workbench for reviewing and navigating the generated operating model, giving consultants a way to inspect AI-generated findings rather than treating model output as an opaque answer.
• Implemented model routing, prompt optimisation, guardrails and PII filtering to make LLM behaviour more reliable and appropriate for enterprise use rather than relying on a single unconstrained prompt/model path.
• Built evaluation and observability into the AI workflows, using traces and repeatable evaluation scenarios to diagnose prompt/model failures and improve behaviour without relying solely on manual testing.
• Worked directly with product and business stakeholders in a highly ambiguous 0-to-1 environment, shaping features and technical approaches as the product evolved rather than implementing against finished specifications.

Built a production-grade AI pilot for a US-based investment network, enabling conversational search and discovery across private CRM data on companies, people, and their relationships. The product combined a React application with an LLM-driven query layer over PostgreSQL, allowing users to ask natural-language questions and receive filtered results, structured answers, and rich result cards.

• Designed and implemented a semantic query-routing architecture for roughly 8–10 distinct search patterns, using semantic classification to send each request through a specialised prompt, tool set and execution path instead of relying on a single generic agent;
• Built LLM tool-calling workflows that extracted structured query parameters from natural-language requests, selected the appropriate database-search tool, executed controlled pre-defined PostgreSQL queries, and used the LLM to rank, filter and format the returned results;
• Used Instructor/Pydantic structured outputs to constrain LLM responses into application-level schemas, allowing the frontend to reliably render structured answers and contextual result cards instead of parsing free-form model output;
• Owned the evaluation, design and implementation of AI safety controls, integrating for input/output guardrails and Microsoft Presidio for detection and filtering of personally identifiable information;
• Implemented end-to-end OpenTelemetry instrumentation and Langfuse tracing across the AI workflows, making model calls, tool execution and request behaviour observable and substantially easier to diagnose during development and pilot use;
• Contributed across the full stack in a four-person engineering team, building features in both the React/TypeScript frontend and Python/FastAPI backend while also owning several of the cross-cutting production-readiness capabilities;
• Helped take the product beyond a disposable PoC architecture, deploying separate development and production environments with automated CI/CD, infrastructure as code, DNS, load-balanced application services, managed PostgreSQL, strict AWS IAM controls and production observability;
• Delivered the pilot successfully enough to lead to a larger production engagement, validating both the product concept and the underlying AI architecture without requiring unrestricted LLM access to the database.

Built an internal customer review intelligence tool for a large B2C software product, providing product teams with deeper insights into Microsoft Store feedback beyond the platform’s native analytics.

• Designed and delivered the application end-to-end in roughly two months, covering product discovery, architecture, frontend, backend, data ingestion, AI analysis, deployment and production operation;
• Built an automated review-ingestion pipeline that periodically collected Microsoft Store reviews through a store endpoint/integration and maintained an internal dataset for historical analysis;
• Implemented LLM-assisted review analysis using low-cost model calls to classify feedback by topic and complaint category and analyse sentiment/tone, turning large volumes of free-text reviews into structured product signals;
• Built custom analytics and visualisations in React/TypeScript to expose review trends, recurring complaints, keyword patterns and changes over time beyond what was available through the Microsoft Store;
• Translated customer feedback into actionable product intelligence, helping the product team identify recurring pain points and make better-informed decisions about which customer issues and product improvements to prioritise;
• Kept the architecture deliberately lightweight for a focused internal tool, avoiding unnecessary platform complexity while still providing automated ingestion, analysis, and a usable deployed product.

Led a multi-year modernisation and migration of a core PDF SDK/platform used across the company’s product ecosystem, with the goal of reducing platform cost, simplifying the technology base and improving the long-term maintainability of the product stack.

• Led technology strategy, architecture, and engineering delivery across multiple products and SDK/platform teams while remaining hands-on with architecture and technical direction;
• Led a multi-year modernization and migration of the company’s core PDF SDK/platform, balancing product continuity, migration risk, and long-term platform simplification;
• Drove the transition from a costly legacy platform dependency to a stronger internal SDK foundation, reducing reliance on third-party technology and increasing engineering control over the platform;
• Coordinated across engineering, product, and leadership teams to sequence the migration, resolve architectural trade-offs, and maintain delivery of customer-facing products throughout the transition;
• Delivered the strengthened platform in May 2024, contributing to approximately 20% annual revenue savings through reduced platform and licensing costs;
• Positioned the organization for AI adoption by improving internal processes, engineering practices, and data-platform capabilities;
• Mentored engineering leaders and developers while remaining closely involved in architecture, product direction, and delivery trade-offs.