Technical Writing
Architecture decisions, agent design patterns, and what I've learned building AI systems and HFT infrastructure.
A strategy at Enigma Securities posted 1.4 Sharpe in backtesting. Live, it was barely profitable. The culprit: 3.2bps spread plus 1.8bps market impact per side — meaning we needed 10bps daily alpha just to break even. Here's the full breakdown of slippage, order flow toxicity, and queue position.
Most crypto bot tutorials end when the code runs. I built HFT infra at Enigma Securities — here's the 5-layer architecture that actually works with real money: feed handler, OMS, risk engine, kill switch, and Cloud Run deployment that doesn't fail at 2 AM.
At Enigma Securities, three engineers built a full HFT stack from scratch: five venues, co-location, sub-15ms tick-to-order, and an automated kill switch that triggered twice and saved both times. Here's every layer of a prop firm's stack and what breaks first when you get any of them wrong.
I ran the same strategy through two backtesting engines at Enigma Securities. One showed 1.8 Sharpe, the other 0.4. Same data. Here's the infrastructure design that eliminates fictitious alpha before you deploy capital.
I built an HFT platform from scratch at Enigma Securities: market data connector, OMS, risk engine, execution layer. Sub-15ms tick-to-order. This is the architecture that hit that target, and the design decisions that actually matter under production load.
I collapsed a 5-agent lead gen pipeline to 2 agents: 22s per prospect down to 8s, 60% cheaper. In 3 weeks it sourced 1,840 contacts, qualified 520, and booked 8 meetings. Here is the architecture, what broke in version 1, and what the system actually costs to run.
At 3 Paris locations, the Shawar'Mama owner hired a manager. At 6, he built an AI agent instead: €28/month, 24/7 uptime, revenue across all locations at 8 AM. Here's the 40% of management time that AI handles better than a person, and when to build it before your next hire.
The Shawar'Mama owner was checking 3 apps for 20 minutes every morning just to understand how yesterday went. Now he reads one Telegram message at 8 AM and moves on in 45 seconds. Here's how the automated AI digest works and how to build one.
The Shawar'Mama owner typed 'best-selling item last Saturday?' mid-investor meeting. Eleven seconds later he had the answer from Zelty, live, without leaving the conversation. Here's the full integration: POS API client, Claude tool use, natural language date resolution, and morning digest.
The Shawar'Mama owner typed a question mid-investor meeting and had his revenue split across 6 Paris locations in 9 seconds — pulled live from Zelty, written in French. Here's the full system: POS integration, automated morning digest, and natural language Q&A on Telegram.
95 negative reviews across 6 Paris locations in the first week. Every one responded to within 5 minutes — without the owner touching a dashboard. Here's the system: GMB API polling, Claude AI drafts, and one-tap Telegram approval.
The ShawaMama review bot handles 95+ reviews a week on a single FastAPI endpoint — returning Telegram webhooks in under 200ms while Claude runs in the background. Here's the full structure: routing, async tool calls, streaming, and Cloud Run deployment.
A LangChain patch upgrade silently changed my client's retrieval pipeline. Chunk sizes shifted, answers degraded, no error thrown. I rewrote the whole thing in 90 lines of Python: chunking, pgvector, OpenAI embeddings, and Claude generation. Here's the complete code.
I built a RAG pipeline for the ShawaMama ops bot, then ripped it out three weeks in and replaced it with two tool calls. The owner asked what sold best yesterday. I had been answering with a 24-hour-old vector index. Here is the decision framework I now run before writing any retrieval code.
The ShawaMama ops bot sends the owner a Telegram message at 8 AM with yesterday's revenue across 6 locations. When a negative review lands, another message fires with a draft reply and one-tap Post. No dashboard, no login, no frontend. Here's why I build every client agent on Telegram.
The ShawaMama bot handles 95-plus negative reviews a week and has run without a restart since I deployed it. It is a FastAPI app calling Claude, living on Cloud Run. Here is the Dockerfile, secrets setup, concurrency config, and webhook pattern I actually use.
Three agents, three copies of the same tool definitions. MCP fixes that: define tools once in a server, and every client discovers them automatically. Here is the complete setup: FastMCP, Claude Desktop config, and wiring it into production agent code.
My first lead gen system had five agents. The rewrite had two. Same output, 60% lower cost, 22s down to 8s per batch. Here is when multi-agent actually helps and when it just adds failure modes.
Most RAG tutorials start with a framework. After one too many silent breakages from LangChain upgrades, I rewrote the core in 120 lines of Python. Here's the architecture that's been running in production.