AI, data and software projects we've shipped

AI assistants, copilots and document automation, plus the data pipelines, APIs and apps they run on. Each project lists the brief, what we built and the stack.

11 case studies
18 verified client reviews on Upwork
Member of Claude Startups by Anthropic

Book a free 30-minute call Browse projects

Clients include

SME Careers
Alvascience
AppLand
CrossroadsCX
Inbound Square
Massdriver
SuperAnnotate

SME Careers

Client

LLM training pipeline — code review

Requirement

Review and harden SME Careers' Python and Docker LLM training pipeline — improve data loading, checkpointing, failure recovery, and operational runbooks so containerized training jobs are production-ready.

What we built

Full audit of training code: batching strategy, gradient accumulation, memory usage, and distributed training configuration
Checkpoint strategy with versioned model artifacts, automatic resume from last good state, and S3-compatible storage layout
Docker image hardening: multi-stage builds, pinned dependencies, GPU runtime flags, and health checks for orchestrators
Runbooks for retraining schedules, hyperparameter sweeps, rollback procedures, and on-call escalation paths
Recommendations for CI validation, dataset versioning with Git, and monitoring training loss and GPU utilization

Stack

Book a call about a project like this

CrossroadsCX

Client

GCP Dataflow — Confluent to BigQuery

Requirement

Stream Confluent Kafka events into BigQuery for CrossroadsCX in near real time, with schema validation, dead-letter handling for bad records, and dashboards for throughput, lag, and error rates.

What we built

Apache Beam pipeline on GCP Dataflow with autoscaling workers and configurable parallelism per topic
Confluent schema registry integration with Avro/JSON deserialization and explicit handling of schema evolution
Dead-letter topic and BigQuery error tables so poison messages never block the main stream
Cloud Monitoring dashboards and alerting on consumer lag, DLQ volume, and pipeline failure rates
Documented deployment, backfill procedures, and cost tuning guidance for ops handover

Stack

Book a call about a project like this

AppLand Inc

Client

REST API — partner integrations

Requirement

Expose a partner-facing REST API for AppLand Inc with OpenAPI documentation, OAuth authentication, rate limiting, consistent error responses, and SDK examples so external integrators can onboard without engineering support.

What we built

Spring Boot REST API with OpenAPI 3 specification, interactive Swagger docs, and OAuth 2.0 client-credentials flows for partner apps
Per-client rate limits, idempotency keys on write endpoints, and structured JSON error codes with actionable messages
PostgreSQL-backed token store, request audit logging, and role scopes so partners access only permitted resources
Sample SDK snippets, sandbox credentials, and step-by-step partner onboarding runbooks
Docker images and deployment configs for staging and production with health checks and rolling updates

Stack

Book a call about a project like this

Millcreek Custom Cabinets Inc.

Client

CabinetSense — SaaS platform & RAG assistant

Requirement

Modernize Millcreek support operations with a Kanban and list ticketing board, embed an AI assistant inside SketchUp for cabinet designers, and deploy production RAG over CabinetSense product documentation with usage-based billing and nightly analytics export.

What we built

Refactored the legacy support board from JSP to lazy JSON APIs with batch Google Cloud Datastore loads, cutting page load time and enabling real-time ticket updates
Built a FastAPI RAG service on Cloud Run with Pinecone retrieval, Cohere rerank, MMR diversity, and multi-LLM routing so answers stay grounded in product docs
Integrated JWT authentication, FastSpring licensing, and usage metering so AI features bill correctly per customer tier
Shipped nightly BigQuery exports for AI usage, feedback scores, and support metrics so product and ops can track adoption
Connected SketchUp extension workflows so designers query documentation and support context without leaving the design tool

Stack

Book a call about a project like this

Finer Technologies, Inc.

Client

Internal copilot — LLM with tools

Requirement

Build an internal copilot for Finer Technologies that answers from Confluence and Notion, runs approved Jira and Slack actions on behalf of users, and enforces role-based access with full audit logging for compliance.

What we built

RAG retrieval layer over Confluence and Notion with chunking, metadata filters, and citation links in every answer
Function-calling orchestration with tool adapters for Jira (create/update tickets) and Slack (post summaries) using scoped OAuth credentials
Role-based access controls so users only retrieve docs and run actions they are permitted to perform
Prompt guardrails, PII redaction in logs, and request audit trails for security and compliance review
FastAPI service with streaming responses, conversation history, and admin dashboard for prompt and tool configuration

Stack

Book a call about a project like this

Komal HK

Client

RAG knowledge base — internal docs search

Requirement

Deliver an internal RAG search system for Komal HK over B2B SaaS documentation — natural-language queries, semantic retrieval from wikis and PDFs, cited answers, and re-indexing when content changes.

What we built

Offline ingestion pipeline: parse PDFs and HTML, chunk with metadata (title, section, URL), embed with OpenAI, upsert into Pinecone
FastAPI search service: embed questions, retrieve top-k chunks, rerank for relevance, and generate answers with inline source citations
Admin endpoints to trigger full or incremental re-index, compare retrieval scores on test questions, and monitor index freshness
Authentication on search APIs, rate limits per team, and logging of queries for quality review
Evaluation harness with golden questions so search quality is measured after every documentation update

Stack

Book a call about a project like this

M2i Consulting

Client

AI Invoice Bot — OCR + LLM extraction

Problem

Eliminate manual invoice entry from photos and PDFs for M2i Consulting by building an OCR plus LLM extraction pipeline that validates party names, line items, and totals before data reaches accounting workflows.

What we built

Tesseract and Google Cloud Vision preprocessing pipeline for skew correction, denoising, and text region detection on photographed invoices
LLM structured extraction with JSON schema validation, confidence scoring, and human-review flags for low-quality scans
FastAPI endpoints for batch upload, async processing queues, and structured error responses when OCR or parsing fails
Dockerized workers with retry logic, dead-letter handling for malformed files, and reconciliation reports for finance teams
Integration hooks so extracted fields map cleanly into downstream accounting and ERP import formats

Stack

Book a call about a project like this

Diana Fernandez

Client

ML feature pipeline — training & serving

Problem

Build a single orchestrated path for Diana Fernandez from raw data through feature engineering, model training, registry promotion, and a production serving API — ending manual per-run feature rebuilds.

What we built

Feature store patterns with Airflow DAGs for daily feature computation and backfill on schema changes
MLflow integration for experiment tracking, model registry, and promotion gates before production deploy
REST serving layer on Kubernetes with autoscaling, latency monitoring, and one-click rollback to prior model versions
Data validation steps between stages so bad upstream data does not poison training or serving
Documentation and handover for the client team to operate retraining and releases independently

Stack

Book a call about a project like this

Employed at KPMG

Consultant · Mar 2020 – Jun 2020

Data governance & quality platform

Problem

Give KPMG data stewards automated profiling, null and correlation checks, and governance workflows at scale — surface quality issues before downstream reports fail.

What we built

Enterprise Data Insight and Data Quality platform for steward workflows, issue tracking, and remediation status
Scalable analytics on Spark and Spark-SQL with Scala for profiling millions of rows across analytical datasets
Automated checks: null rates, distribution shifts, correlation anomalies, and cross-table consistency rules
Contributed to data modeling and architecture sessions, improving pipeline design and ownership boundaries
Dashboards for stewards showing trend lines, open issues, and SLA compliance per domain

Stack

Book a call about a project like this

Polina H.

Client

Full-stack dashboard — pipeline health

Problem

Give Polina H. one dashboard for pipeline health instead of checking Airflow, logs, and email separately — SLA status, run history, and quality-check alerts in a single view.

What we built

Flask backend aggregating Airflow run metadata, task durations, and data quality check results into a unified API
React and TypeScript frontend with Tailwind: status cards, filters by pipeline and date, and drill-down detail panels
Alert hooks integrated with existing Slack and email notification channels when SLAs breach or checks fail
Role-based views so ops sees all pipelines while team leads see their owned workflows
Deployed on GCP with authenticated access and documented refresh intervals for each data source

Stack

Book a call about a project like this

Shenali De Silva

Client

Aadhya Self Drive — Flutter iOS & Android

Requirement

Ship a cross-platform mobile app for Aadhya Self Drive so renters browse vehicles, complete verification, book trips, and pay on iOS and Android from one Flutter codebase, with App Store and Play Store release.

What we built

Flutter app for iOS and Android with location-based search, flexible rental durations, and end-to-end trip management
Booking flow with document upload, renter verification steps, and payment integration tied to the rental backend
TestFlight and Google Play release pipeline including store listings, screenshots, and versioned rollout updates
Push notifications for booking confirmations, trip reminders, and status changes
Performance tuning for offline-tolerant flows, deep links from marketing pages, and analytics on conversion steps

Stack

Book a call about a project like this

Stack