Max – React, Python, FastAPI

An automated SaaS solution that manages and streamlines utility billing for property management companies, including an AI assistant that answers questions over billing documents and resident records.

• Joined the project from scratch and delivered it end to end — API design, data layer, UI and deployment — as a full-stack engineer on a small team.
• Built the assistant's retrieval over billing documents and resident records on PostgreSQL + pgvector — structure-aware chunking that keeps charge tables intact, and hybrid vector + keyword search with re-ranker after pure vector search proved unreliable on exact identifiers such as charge codes and invoice numbers.
• Added retrieval guardrails — tenant-scoped metadata filters, deduplication, a context cap and a relevance threshold that returns an explicit "no relevant context" answer instead of a fabricated one.
• Designed query routing in LangGraph — exact values resolved through SQL via Pydantic-validated tool calls, open questions through retrieval — with every answer returned as schema-validated JSON.
• Cut response latency and token cost through context trimming, tiktoken accounting and model routing, and measured retrieval accuracy in LangSmith to keep changes evidence-based.
• Integrated third-party provider APIs, including ResMan, into a single canonical schema.
• Worked with high-volume transactional data: redesigned data models to eliminate large-scale duplication and accelerate billing calculations, and tuned queries with EXPLAIN ANALYZE, covering indexes and Redis caching.
• Split storage by access pattern — meter readings, a high-write append-only time series, in Apache Cassandra, partitioned by meter and billing period so a billing run reads one partition per meter; charges, invoices and resident records in PostgreSQL, where billing needs joins and transactions.
• Built the property-manager interface in Next.js and Material UI — billing and resident views and XLSX billing exports for downstream reporting.
• Owned delivery: CI/CD via Bitbucket Pipelines, full Docker-based deployment on AWS EC2 behind Nginx, plus continuous refactoring, testing and bug-fixing.

An admin and submission platform for an annual field-data collection program — coordinating geographic survey areas, annual events and participants, and verifying observation submissions across regions.

• Built the FastAPI backend around geospatial survey areas from the project's PostGIS/GeoAlchemy2 schema and Alembic migrations, Shapely geometry validation and spatial queries.
• Extracted typed fields from free-text submission notes using structured outputs with Pydantic schemas, constraining transform prompts to rewriting existing content without introducing new facts — essential because the records feed a scientific dataset.
• Built LLM-assisted verification that pre-classifies field observations against an external taxonomy and flags anomalies, using confidence thresholds and human-in-the-loop review instead of automatic rejection.
• Built retrieval over program documentation and historical records on PostgreSQL + pgvector as a LangGraph retrieve → validate → summarize flow, with role and region filters applied at retrieval time so multi-tenant permissions hold inside the AI features.
• Benchmarked hosted models (Anthropic, Gemini) against self-hosted Ollama models (Mistral, Phi-3) on quality, latency, cost and data privacy, and recommended the setup the program adopted.
• Built a BigQuery analytical layer for cross-region, multi-year aggregations, using partitioning and clustering to keep reporting queries fast and predictable in cost.
• Designed the multi-tenant role and permission model, enabling administrators to delegate coordinator rights across regions.
• Built photo submission through S3 presigned uploads with HEIC normalisation, so field photos went from the device straight to storage without passing through the API.
• Built the Next.js App Router page and layout hierarchy and implemented forms with React Hook Form + Zod validation.
• Covered CRUD, workflows and error paths with Pytest, deployed via AWS CloudFormation to ECS Fargate, and contributed to refactoring and code reviews in a Scrum team.

A multi-provider rail booking back end that aggregates train schedules, fares, seat selection and ticket issuance from several third-party rail providers into a unified booking API.

• Built the unified search API over rail providers — parallel async fan-out with partial results, so one slow provider couldn't stall a search.
• Implemented the booking state-machine flow and its orchestration across independent providers, and designed the cross-provider search aggregation.
• Built refund and cancellation workflows with provider-specific policy calculation.
• Published booking events to Kafka for downstream services with delivery guarantees.
• Integrated orders and payments through an internal PCI-DSS tokenization service.
• Implemented a Redis-backed station autocomplete index and S3 ticket-asset storage with secure download.
• Built the test harness around Mountebank-stubbed rail providers, so booking flows ran end to end in CI without live provider sandboxes.
• Contributed to bug-fixing, refactoring and code reviews in a Scrum environment.

A CRM that simplifies the sales process: managing customer information, tracking results, creating tasks and calculating statistics.

• Built record-level ownership and role-based access, so each manager saw and edited only their own customers and tasks.
• Implemented automated regular follow-ups for potential clients, reducing manual tracking for the sales team.
• Built the CRM's DRF API and its Vue.js front end (Vuex, Vue Router) — customer, task and pipeline-statistics views, with tag- and comment-based filtering for faster pipeline analysis.
• Managed infrastructure across AWS, Docker and GitHub Actions, and kept code and database optimized and maintainable.
• Resolved front-end and back-end defects from QA, covered code with unit tests and ran development testing.
• Took part in onboarding new developers and collaborated in a Scrum environment.
• Integrated Stripe invoicing into the sales flow — managers issued invoices from a customer record, and payment status synced back through webhooks to mark deals as paid in the pipeline statistics.

A platform that lets teachers and students run group online lessons and communicate in chats.

• Built email invitation management and delivery for classes and sessions.
• Enabled shared note-taking during live calls, giving students a collaborative record of each lesson.
• Developed the API together with the matching front-end functionality, including React components and page design.
• Set up Docker environments and CircleCI pipelines.
• Improved the testing pipeline and expanded unit test coverage across key components.