Varsha Thondalapally

Applied AI Engineer · Employee #1 product builder · Austin, Texas

LinkedIn · GitHub

CONFIDENTIALITY NOTE — This case study describes work performed as an employee. The product, source code, design, data, and intellectual property belong to my former employer. No company code, screenshots, prompts, customer data, credentials, schemas, or vendor payloads are reproduced here.

60-second view

I am an Applied AI Engineer who works on the difficult layer after the model demo: turning probabilistic output into product behavior that is exact where it must be, reviewable by people, durable across retries and restarts, and safe at billing and external-system boundaries.

At an early-stage vertical-AI company, I served as employee #1 and the only full-time engineer. Starting from a minimal React and FastAPI scaffold and a company-supplied concept, I developed substantial parts of a company-owned workflow from an approximately five-week working vertical slice through deployed production operations.

The system combined multimodal evidence, retrieval, structured model outputs, deterministic validation, human review, durable workflow state, billing, third-party delivery, and AWS operations. The central rule was simple: AI could propose; authoritative data, current state, commercial authorization, and a human decision controlled what happened next.

At a glance

• Role: Employee #1 — Applied AI & Product Engineering
• Inputs: documents, pages, photographs, image crops, sketches, measurements, notes, structured intake, and audio
• AI layer: OpenAI, Gemini, vision, embeddings, FAISS, hybrid retrieval, reranking, structured outputs, and transcription
• Product layer: React, TypeScript, Python, FastAPI, Pydantic, PostgreSQL, SQLAlchemy, Alembic, S3, Redis, SQS, and Lambda
• Business boundaries: Stripe lifecycle, entitlements, usage identity, and third-party estimating-platform delivery
• Operations: Docker and AWS deployment, monitoring, security, migrations, releases, diagnostics, and recovery

What changed from idea to operated product

The product evolved because the risks became clearer as the workflow became real.

1. Vertical MVP: authentication, protected routes, clients, sessions, database migrations, multi-file upload, assessment persistence, grounded retrieval, AI follow-up questions, editable output, and assignment state.
2. Room-aware workflow: evidence became associated with rooms, pages, sources, measurements, sketches, and estimate state rather than treated as one undifferentiated prompt.
3. Conversational workflow: conversation helped operators clarify ambiguous evidence, while structured controls retained exact business fields and state.