Agents you can trust with real systems.

Work · 01LangGraph · FastAPI MCP · Python

Autonomous campaign agent team

A marketing team implemented as four LangGraph agents with one job each: Research pulls live market data through web search and returns structured JSON, Strategy turns it into positioning, Content drafts for blog, LinkedIn and Twitter, and the Reviewer grades each draft PASS or FAIL.

The edge I care about is the FAIL route: the graph sends the draft back to Content with the reviewer's notes attached and iterates until it passes. The routing decision is made in code from the verdict field, not left to the model. Tools live outside the agents in a FastAPI MCP server, so a tool can be swapped or mocked without touching agent logic.

• StateGraph
• conditional edges
• shared state
• web_search · Tavily
• write_file

View the source

Work · 03spaCy · FAISS + MiniLM · Gemini

RAG over financial news, two ways at once

A research assistant over forex news that runs two retrieval paths and compares them. Path one is a knowledge graph: spaCy named-entity recognition and a dependency parse over every article, extracting entities and subject-verb-object relations. Path two is dense retrieval: MiniLM sentence embeddings in a FAISS IVF index, inner-product search, top ten chunks per query.

Both contexts go to Gemini separately and the stronger answer wins. Around that sit an LRU cache on queries, a SHA-256 change tracker so unchanged documents are never re-embedded, a news fetcher that writes timestamped snapshots, and three MetaTrader 5 signal scanners whose output joins the same corpus.

• NER + SVO graph
• FAISS IVF · top-10
• SHA-256 change tracking
• MT5 EMA / bias / trend

View the source

Case · 01Cross-border payments fintech · current role

A rate engine that picks the payout partner

A cross-border payment can usually go out through several payout partners, and each one quotes its own rate for each corridor and each way the money is paid out. Choosing well means comparing quotes that arrive in different shapes from different providers.

I built the product that does it end to end. It fetches rates from the providers, normalises them into one shape, and for every corridor and disbursement mode returns the best rate and the partner that offers it. The employer stays unnamed here and there are no numbers, but the shape is the same one I use everywhere: every provider sits behind one contract, so adding a partner does not touch the ranking logic.

• multi-provider fetch
• normalised quotes
• corridor × payout mode
• best rate + partner

Case · 03My own system · Python · IMAP/SMTP · Apps Script · Playwright

An outreach engine that runs with the laptop off

I needed to find businesses, enrich them with a real contact, write to each one individually and follow up, without a paid sending tool and without burning sender reputation.

The engine discovers leads (Google Places plus a crawl of each site for published emails and an MX check), keeps them in a merge-safe store, and schedules every message into the Drafts folder of one of 14 Gmail accounts over IMAP. A small Apps Script outbox on Google's servers sends each draft at its time, so nothing depends on my machine. A reconcile step reads Sent Mail back into the CRM, and a scanner picks up replies and bounces across every inbox.

The useful part was what broke. A seed test showed three of the accounts were landing in Spam, so they were taken out of rotation. Two runs writing the store at once lost rows, so saves now merge. A failed upload that had actually succeeded left an orphan draft that would have sent twice, so every queue is now checked message by message against Drafts.

• 14 sender accounts
• cloud outbox
• per-account caps
• reconcile + reply scan
• seed placement tests