August 2026

The problem

Qualfix needed one internal management and CRM system spanning two legal entities (Qualfix BD and Qualfix USA), including multi-currency operation.

What I built

Deployed and configured Odoo 19 Community on self-hosted Docker infrastructure via Dokploy, covering the CRM → Sales → Project → Invoicing pipeline, custom branded modules (login theme, report layout, portal theming), multi-currency support and banking integration.

Outcome

Both entities run day-to-day operations on it, maintained in-house rather than outsourced.

July 2026

The problem

Fourteen years of historical business data sat in Salesforce and had to move into Odoo, where the schema and structure did not line up field-for-field.

What I built

An AI agent-assisted migration: agents reconciled schema and structural differences between the two systems, with the mapping reviewed rather than hand-written record by record.

Outcome

Completed in 4 weeks. Work that would typically take months of planning.

April to May 2026

The problem

70-100 sales emails arrived daily carrying PDFs, images and spreadsheets in no fixed format. Extracting them into a standard template was entirely manual.

What I built

An agent built in LangGraph classifies each email and extracts structured records in a single Gemini 2.5 Pro call, with the output schema enforced by Pydantic so nothing downstream parses free text. MLflow 3.x traces every email with token cost; Prometheus and Grafana track queue depth, p95 latency and monthly billing. Runs on FastAPI and async PostgreSQL in Docker Compose.

Outcome

Daily processing dropped to roughly 5 minutes, down from a fully manual process, and the cost and latency of every run is now visible.

March 2026

The problem

Tier 1 technical support calls follow deterministic workflows, but callers rarely stay on script, so a rigid IVR fails and a purely free-form LLM drifts.

What I built

Combined structured call logic with LLM-driven conversation: the agent follows the defined support workflow while handling dynamic, off-script situations, built on Retell AI with Gemini, OpenAI and ElevenLabs.

Outcome

Proof-of-concept covering inbound and outbound calls; currently in review.

2025

The problem

Group Purchasing Organization contract data had to be matched against end-user records across inconsistent entity names. Anything the system missed fell to manual reconciliation, which did not scale with contract volume.

What I built

Reworked the matching logic and cleaned the underlying data: normalising entity names, resolving structural differences between sources, and tightening the comparison rules rather than reaching for a heavier model.

Outcome

Automatic match rate went from 37% to about 88%, removing most of the manual reconciliation load.

December 2025

The problem

Hiring teams read applications in arrival order, so the strongest candidates are not necessarily the ones seen first.

What I built

A Next.js and PostgreSQL job board with an LLM that scores each applicant against the job description. Background processing and email notifications run on Inngest, keeping the request path fast.

Outcome

Reviewers see ranked candidates rather than a raw queue. Source is public on GitHub.