§00/ The operator Available Q3 2026

I build custom AI systems that replace the work you stopped enjoying.

AI Engineer RAG Agents Python · FastAPI · FastMCP · ADK · Vertex AI
$ rag-agent --session new live
model=gpt-4o vectors=2048d top-k=8 conf >= 0.85
>
simulated session · real reasoning, mock retrieval
§01/ Case study ai-sales-agent
role architect · sole engineer
stack Python · Google ADK · Gemini · Google Search & URL Context
status public demo · premium tier on request
replaces 20 minutes of manual prospect research per lead

A multi-agent orchestrator that researches a prospect and drafts a personalized cold email before your coffee is ready.

Problem

B2B cold outreach lives or dies on personalization. Real personalization takes 15 to 30 minutes per lead: scrape the site, read recent news, infer pain, write a tight email. At scale, teams either skip it and tank reply rates, or burn out doing it and churn SDRs.

Build

Four agents on Google ADK with Gemini. URLReaderAgent scrapes the prospect's site. SearchAgent pulls recent news and announcements. CodeAgent handles any data the first two can't. SalesOrchestratorAgent synthesizes everything into a concise, non-robotic cold email. Default pitch is your AI services. Swap the prompt in app.py to pitch SEO, lead-gen, web dev, anything B2B.

Result

Prospect research collapses from 20 minutes to about 15 seconds. Cold-email reply rates lift because every email is grounded in live context. The orchestrator pattern transfers: same shape works for inbound triage, RFP drafting, support escalation, any workflow that needs live context plus a written artifact. Premium tier (client work) adds CRM hooks, bulk processing, custom knowledge bases, and a human-in-the-loop approval UI.

§02/ How I work six phases · a real SOW · scroll to reveal
01 · Diagnose

20-minute screen-share. I read your stack. We name the three bottlenecks.

I skim your docs, your CRM, your slowest workflow. You leave the call with a one-page note: here is what I would build first, here is what I would not, here is the rough cost. Free. No deck.

02 · Spec

One-page architecture doc. You sign it. We both know what we are shipping.

Inputs, model choice, retrieval design, eval set, latency budget, failure modes, success metric. Tight. Two rounds of edits max. Anything that does not serve the metric gets cut.

03 · Prototype

Working RAG or agent in 5 to 7 days. Real data, not a demo dataset.

FastAPI service exposing REST APIs, pgvector on your infra, custom ETL pipelines pulling your docs and CRM, eval harness running against your labeled examples. You get a URL and a Loom. We iterate weekly.

04 · Harden

Eval set, guardrails, latency p95 under 800 ms. Production-grade or nothing.

Prompt injection tests, refusal behaviour, citation accuracy, retry/backoff, observability, kill switches. The unglamorous work that decides whether the system actually survives Monday morning.

05 · Ship

Deploy. Monitor. On-call for two weeks. I do not ghost after launch.

Canary at 5%, watching logs, watching eval drift, paging me when something looks off. Two weeks of post-launch on-call is in every contract, no extra invoice.

06 · Handoff

Runbook. Loom walkthrough. Optional monthly retainer.

Your team owns it. I write the runbook, record the walkthrough, sit on two calls while your team takes over. If you want me on standby, the retainer is fixed, capped, and cancelable.

§03/ Contact

Twenty minutes. Your stack on screen. No deck, no pitch.

If there is a workflow eating hours your team hates, bring it. I will tell you whether an AI system is the right fix or not, on the spot.

Book a free audit cal link · calendly.com/osamatech786-jdqf/20min