M.G.Software · AI Automation

Available worldwide

Software Engineer&AI Automation Specialist.

I build the systems that drive operational efficiency at scale: AI-augmented workflows, automation pipelines, and the software that helps teams handle more volume and grow without adding headcount.

Scroll

Python●TypeScript●LangGraph●FastAPI●n8n●React●Three.js●Docker●Postgres●OpenAI●Claude●GCP●Python●TypeScript●LangGraph●FastAPI●n8n●React●Three.js●Docker●Postgres●OpenAI●Claude●GCP●

01About

Built on curiosity, sharpened by work.

0×

Faster diagnostics

Cut fault-diagnosis time from 18 minutes to 3 with a LangGraph-based AI system for biomedical equipment technicians.

Off the reporting process

Replaced a 15-minute manual spreadsheet update with a Telegram bot that syncs field data to Google Sheets in under a minute.

LLM outputs graded

Evaluation work at Scale AI: reasoning, instruction-following, code quality. Taught me how models actually fail.

Background

Biomedical roots, software career.

I studied biomedical engineering. Most of my time outside class went into software — frontend first, then backend systems, then automation workflows, then AI tooling.

The work I keep coming back to lives at the edge of engineering and operations. Finding where teams lose capacity or leave revenue on the table, then building the systems that recover both.

What I build

Internal tools
AI-augmented workflows
Automation systems
Modern web applications

02Selected work

Selected work, real results.

Email infrastructure, diagnostic systems, appointment platforms, field automation. Each one cut real cost off a real workflow.

01HR & Staffing · Email AI2026

AI Email Infrastructure

Specific Talent Ltd is an HR and staffing company supporting brands like Quickmart, Carrefour, Decathlon, and Pharmaplus. Their team handled high email volume where genuinely urgent messages were easy to lose inside a manually-sorted inbox. I upgraded their email infrastructure so the conversations that matter surface first, and routine handling no longer drains the team's time.

Result

The team identifies priority communications faster, handles far less manual inbox triage, and keeps response cycles tight on the conversations that move the business.

Priority communications surface first, so the team acts on what matters before it goes cold.
Manual inbox handling cut down across high-volume daily email.
Faster, more consistent response cycles on critical business conversations.

Custom AI rules01 / 02

HR & Staffing · Email AI

02Healthcare · LangGraph2026

AI-Augmented Troubleshooting System

Biomedical equipment faults used to stall junior technicians for up to 18 minutes because the diagnostic path lived inside the heads of senior staff. This system externalises that knowledge. LangGraph structures the diagnostic flow, an LLM interprets symptoms and suggests next steps, and a small equipment ontology keeps the reasoning grounded in real fault patterns.

Result

Average fault-diagnosis time dropped from 18 minutes to 3. Most faults no longer require a senior technician in the room.

LangGraph state machine routes between diagnostic stages without hallucinating paths.
Equipment ontology grounds the LLM — no generic medical advice, only known fault patterns.
Operator-in-the-loop at every step. The technician confirms or overrides before anything runs.
Full audit trail of every decision, retained for compliance review.

PythonLangGraphLangChainOpenAIFastAPIPostgreSQL

diagnostic-pipeline · activelive

Symptom intake0.2s
Fault classification0.8s
LLM diagnosis1.4s
Operator confirm--
Action plan--

llm output · streaming

Fault pattern matches: power-supply undervoltage on Module B.Recommended: check rail voltage at TP7. Expected 12V, likely 9–10V.

Before18 min

After3 min

Speed-up6×

03Healthcare · Web platform2024

MedAppoint

Medical appointment platform built for clinics and patients in Kenya. I owned the patient-side booking interface: the flows, the responsiveness, and the parts that break when the network does. The challenge was making a booking experience that actually works on mid-range Android phones on 3G.

Result

Booking flow held up on 2G. Page load time dropped by cutting unnecessary re-renders and tightening component performance.

Eliminated redundant re-renders with careful state isolation and memoisation.
Loading states and skeletons tuned for slow connections, not fast ones.
Simplified the booking flow so patients could confirm appointments in fewer steps.
Keyboard navigation and screen-reader labels throughout.

ReactTypeScriptRESTTailwindVite

Home01 / 06

MedAppoint · UI

04Climate-tech · Automation2025

Field Reporting Automation

At an electric mobility startup, field operators sent financial data over WhatsApp and the COO manually copied it into a spreadsheet. Every day, 15 minutes gone. I automated the whole path (Telegram intake, validation, Google Sheets sync) and made it reliable enough that the COO stopped checking the work.

Result

15 minutes of daily manual work replaced by a bot that runs in under a minute. Zero double-writes in the first 30 days of operation.

Validation on intake catches malformed entries before they reach the spreadsheet.
Retries and idempotency on every external call — flaky mobile networks don't cause duplicate records.
Telegram interface, because that's the tool the field team already had on their phones.
Reconciliation logic keeps the spreadsheet consistent even when ops run out of order.

n8nNode.jsTelegram APIGoogle Sheets APIDocker

finance-automation · runninglive

✈

field operator

⟳

Validate

format · range · dedup

▦

Sheets sync

idempotent write

execution log · last 30d

09:14:03✓Intake validated · operator @james_m

09:14:04✓Idempotency check passed · writing row 847

09:14:04✓Sheets sync complete · 0 duplicates

Before15 min

After< 1 min

Errors / 30d0

03Tools

The tools I actually use.

Grouped by what I use them for. Not exhaustive: these are the ones I reach for most weeks.

AI Engineering

Agents and LLM workflows built to hold up in production. Most of my AI work is here.

LangGraph92

LangChain90

OpenAI APIs95

Claude APIs92

Evals & Calibration88

04Experience

Where I've worked.

Short list, four teams, four very different setups. A few notes on what the work actually was.

2025
E-MOTI
Climate-tech · Electric mobility
AI Automation Engineer
Built the financial reporting automation from scratch. Stack was n8n, a Telegram bot, and a few Google APIs. The constraint that mattered was reliability: finance needed to trust the output, so every call had retries, idempotency, and validation before anything hit the spreadsheet.
- 15-minute daily manual process down to under a minute
- Zero duplicate records in the first 30 days of operation
- Field team didn't need to change tools — Telegram was already on their phones
2024
MedAppoint
Healthtech
Frontend Engineer
Built the patient-side booking interface in React. Most of the work was performance: making the app usable on the devices and networks Kenyan clinics actually run on. The rest was standard API integration, UX iteration with the product team, and accessibility.
- Booking flow stayed functional on 2G connections
- Keyboard navigation and screen-reader support throughout
- Component patterns the team adopted for the rest of the product
2024
Petshelpful
Consumer web
DevOps Engineer
DevOps-focused role on a small consumer product team. Owned the deployment infrastructure and release pipeline that kept the engineering team shipping reliably.
- Managed CI/CD pipeline and deployment workflows end-to-end
- Maintained environment consistency across dev, staging, and production
- Reduced release friction so the product team could ship faster
2023
Scale AI / Remotasks
AI data · Evaluation
AI Quality Specialist
Graded LLM outputs across reasoning, code, and instruction-following. The most useful background I have for building anything with LLMs. You stop expecting models to be smart and start watching how they fail.
- 800+ outputs evaluated across reasoning, code, and instruction tasks
- Developed a pattern library for hallucination and reasoning failures
- Feedback fed directly into production model training runs