Nairobi, Kenya · Open to remote worldwideAvailable for work

Gevverson
Kipkoech

AI Software Engineer · Full-Stack Engineer · LLM & Agent Evaluation

I design and ship production-grade products end-to-end — AI/RAG platforms, mobile healthtech, and payments/fintech systems — from first commit to live deployment.

View WorkGet in touchDownload Resume

Featured Work

Things I've shipped

Production-grade products across AI, healthtech, civic tech, and fintech — from database schema to live deployment.

Macho ya Mwananchi screenshot

Civic Tech / AI

Macho ya Mwananchi

LIVESolo build

A civic-accountability platform exposing ghost projects and misused public funds in Kenyan counties — built on official Auditor-General audit data.

The Problem

Government audit reports are public but locked in 400-page PDFs ordinary citizens can't read.

Ksh 3.3B

Funds tracked

144

Facilities closed

1,885

Ghost workers

Auditor-General

Data source

  • AI Audit Translator — ask questions in Swahili or English via a RAG pipeline (ChromaDB + Google Gemini) with a direct-Gemini fallback
  • GPS citizen verification: photo + live coordinates feed a consensus engine that updates a project's status
  • Anonymous corruption reporting with tracking codes
  • Interactive Leaflet evidence map with real government data overlays
React 18ViteTailwindLeafletFramer MotionNode/ExpressMongoDB AtlasFastAPIChromaDBGoogle GeminiDockerGitHub Actions
CodeLive Site
Helix screenshot

Healthtech · Full-Stack + Mobile

Helix

Case StudySolo build

A genetic-compatibility dating platform: users store an encrypted genotype/blood profile (typed or OCR-scanned from a medical report), then match with a sickle-cell / rhesus "traffic-light" verdict that gates whether matched profile photos un-blur.

The Problem

Couples need accessible sickle-cell & blood-compatibility guidance before starting a family — without exposing deeply sensitive genetic data.

27

API endpoints

~150

Backend tests

15

Genotype pairs

12

Data models

  • FastAPI + PostgreSQL backend (SQLAlchemy 2.0, 13 Alembic migrations) with a 15-pair genotype engine that explains per-child inheritance risk (e.g. 1-in-4) as a GREEN/AMBER/RED verdict — framed as educational, not diagnostic
  • Field-level Fernet (AES) encryption of genotype/blood data via a custom SQLAlchemy type; JWT + bcrypt + server-verified Google OAuth; per-route rate limiting and an immutable audit log
  • Privacy-by-design: genetic markers pseudonymised behind a random subject-ref decoupled from user identity; GDPR / Kenya-DPA consent gates, an 18+ gate, and a right-to-erasure cascade
  • Flutter app (Provider + go_router auth guards, silent re-auth, FCM push, 16 screens), Tesseract OCR to read a report photo, and server-enforced photo-blur via signed Cloudinary URLs so blur can't be bypassed
FastAPIPostgreSQLSQLAlchemy 2.0AlembicFernet encryptionJWT + Google OAuthTesseract OCRFlutter/DartDocker
APK / demo available on request.
AI Trading Bot screenshot

Fintech · Quant Systems

AI Trading Bot

Open SourceSolo build

A self-hosted algorithmic-trading research harness — a 9-gate signal + risk pipeline (Kalman, HMM regime detection, Bayesian sizing, half-Kelly) over TimescaleDB time-series, with an MT5 demo bridge and Telegram alerts. Paper-trading by default.

The Problem

Learning quant systems design means building the full loop — market data, strategy, multi-stage risk gating, execution, and monitoring — not just a single indicator.

9

Risk gates

7

Quant modules

23

API endpoints

~5.1k LOC

Backend

  • FastAPI + TimescaleDB hypertables + Redis + ChromaDB across 4 Docker services; live market data via Yahoo Finance
  • 9-gate signal & risk pipeline: news blackout, drawdown circuit-breaker, Kalman trend, multi-timeframe confluence, VWAP order-flow, HMM regime, conviction sizing
  • 7 quant-math modules — Kalman filter, 3-state HMM, Bayesian win-probability, half-Kelly sizing with a 5% cap and negative-expectancy rejection
  • Paper-trading journal by default; optional MT5 demo order routing (MQL5 EA) and Telegram signal broadcasting
FastAPITimescaleDBRedisChromaDBMT5 (demo)TelegramDocker ComposeReact/Vite
Personal / educational — paper-trading by default; no live-money or performance claims.Code
Intelliverse — AI Agent Builder screenshot

Agent Builder · AI / LLM Engineering

Intelliverse — AI Agent Builder

LIVEInternship · Team Lead

A deployed multi-user platform where anyone can build custom AI chat agents — each with its own persona and uploaded document knowledge (RAG) — then share or fork them.

The Problem

Non-technical users want AI assistants grounded in their own documents, without writing code or handling an API key themselves.

21

API routes

49

Backend tests

Pinecone

RAG store

Fly.io

Deployed on

  • Flask (app-factory, 4 blueprints) backend with server-rendered Jinja + vanilla-JS frontend; PostgreSQL via SQLAlchemy + Alembic; Dockerised and deployed on Fly.io with a /health readiness check
  • Document-RAG: users upload PDF/DOCX/TXT → 500-char chunking → Google Gemini embeddings → Pinecone vector retrieval (top-k) grounding each agent's answers
  • Hardened auth & safety: Google OAuth + password login with account lockout, CSRF protection, Talisman CSP/HTTPS, per-route rate limiting, and an immutable audit log
  • Resilient LLM calls via a model-fallback chain (Gemini 2.0-flash → 1.5-flash → 1.5-pro); agents and chats can be shared and forked between users

Built during an internship as the lead engineer, working alongside 1–2 other contributors.

PythonFlaskPostgreSQLSQLAlchemyPineconeGoogle GeminiDockerFly.io
CodeLive Site
Payment Orchestration Platform screenshot

Fintech · Payments · Team project

Payment Orchestration Platform

Case StudyTeam · Core contributor

A multi-country payments platform routing mobile money (M-Pesa, MTN MoMo, Airtel Money) and card payments through one provider-agnostic core.

The Problem

Businesses operating across Africa need a single integration that handles mobile money and card payments without per-provider re-engineering.

  • My contribution: mobile-money payment-rail adapters and the hosted checkout flow
  • Webhook verification and idempotent transaction handling
  • Hosted checkout so card data never touches the server

Team project owned by a collaborator; I contributed as a core developer.

TypeScriptNestJSPostgreSQLDocker
Private team repository — details and demo available on request.

Team & client contributions

Private repositories — code not published without the owners' permission.

Clinical Records System

Team · Contributor

Production records system for a cardiology practice — CI quality gates and delivery-process improvements.

Nuxt 3 / VueTypeScriptPostgreSQLPlaywright

Membership Records Platform

Team · Contributor

Multi-tenant membership records platform — access scoping, data validation and shared UI components.

Next.jsTypeScriptDrizzle ORM

Also built

Jirani Alert System

Community-safety platform with role-based admin/responder/community dashboards.

PHPMySQL

Kings Website

Vanilla-JS event-ticketing app with cart, checkout, and unit tests.

JavaScriptNode

Experience

Where I've worked

Terminal-Bench Reviewer (Contract) · Turing

Aug 2026 – Dec 2026 · Remote

  • Review agentic coding-task packages before they enter the benchmark — auditing the instruction, test suite, oracle solution, Dockerfile and run harness for false positives (a wrong solution passing) and false negatives (a correct solution failing).
  • Wrote and iterated a 29-step reviewer protocol through 7 revisions — declared-vs-tested behaviour cross-reference, hidden-gate and harness anti-tamper scans, oracle / no-op gates, determinism and budget checks — every finding tied to quoted evidence.
  • Graded tasks alongside frontier-agent rollouts (pass@8, Claude Code and Codex) across ML-infra and backend domains.

Lead AI Engineer · Helix

2026 – Present · Remote

  • Building a genetic-compatibility platform end-to-end: FastAPI + PostgreSQL backend (27 endpoints, ~150 tests) and a 16-screen Flutter app.
  • Privacy-by-design: field-level encryption of genetic data, pseudonymisation, JWT + Google OAuth, and a right-to-erasure cascade (GDPR / Kenya-DPA).

Lead Developer · Constituency Tech Initiatives

2025 – Present · Nairobi, KE

  • Built Macho ya Mwananchi, a live civic-accountability platform on official Auditor-General data (Ksh 3.3B tracked) with a Swahili / English RAG audit assistant.

Software Engineer (AI & Agents) — Team Lead · IntelliVerse

Nov 2025 – Jan 2026 · Remote

  • Built and deployed a multi-user AI agent-builder on document RAG (Flask, Gemini embeddings, Pinecone) on Fly.io; led a small team (2–3) through sprints and code reviews.

LLM Trainer & Quality Evaluator (Contract) · Turing

Dec 2025 – Feb 2026 · Remote

  • Evaluated frontier LLMs on agentic / computer-using tasks; produced red-teaming reports and prompt-failure taxonomies across experimental checkpoints.

Data Annotation Specialist · iMerit (Ango Hub)

2023 – 2024 · Remote

  • Annotated multi-modal datasets under RLHF protocols, focusing on reasoning chains, code-generation quality and factual grounding.

Data & Software Engineering Intern · Kenya National Bureau of Statistics (KNBS)

2023 · Kenya

  • Built Python (Pandas / NumPy) pipelines to clean and validate demographic datasets, plus Flask + JavaScript dashboards for county-level data.

Education: B.Sc. Statistics & Programming, Kenyatta University (2024) · B.Sc. Computer Science, University of the People (in progress).

About

Engineer who ships

I'm an engineer who ships complete, production-grade products from research to deployment — solo across AI, healthtech, civic tech, and fintech, and as a core contributor on a team-built payment platform.

Comfortable across the full stack: TypeScript/NestJS and Python/FastAPI on the backend, React/Next.js and Flutter on the frontend, and LLM/RAG systems with Pinecone and ChromaDB vector search and Google Gemini at the AI layer. I care about the details that matter — security, performance, and real-world correctness.

Right now I'm contracted to Turing as a Terminal-Bench reviewer — auditing agentic coding-task packages for false positives and false negatives before they enter the benchmark. That evaluation work sharpens how I build: tests that actually prove behaviour, reproducible Docker environments, and evidence over assumptions.

Based in Nairobi, Kenya. Open to remote work worldwide. Whether it's a 0-to-1 product, a complex integration, or an AI feature you need to get right the first time — I'm built for that kind of challenge.

WhatsAppConnect on LinkedInSend an Email

Skills

What I work with

Backend

NestJSNode.jsExpressPythonFastAPIFlaskDjangoCeleryPrismaSQLAlchemyPostgreSQLMongoDBRedisTimescaleDB

Frontend

ReactNext.jsTypeScriptViteTailwind CSSNuxt / VueFlutter/Dart

AI / LLM

RAG PipelinesVector SearchPineconeChromaDBMem0Prompt EngineeringLLM Evaluation & Red-teamingGoogle GeminiOpenAI & Claude APIsLangChainStructured OutputOCR / Tesseract

Evaluation / QA

Terminal-Bench / Harbor Task ReviewFalse-positive & False-negative AnalysisOracle & Test-suite ValidationRubric Designpass@k Agent RolloutsDocker ReproducibilityRLHF Annotation

Payments / Fintech

Payment-rail IntegrationM-Pesa (Daraja)MTN MoMoAirtel MoneyCard TokenizationIdempotencyWebhook Signature VerificationPCI-aligned Design

DevOps

DockerDocker ComposeGitHub ActionsBranch Protection & Quality Gatespnpm MonoreposFly.ioRenderVercelLinuxNginx

Security

JWTOAuth2Field-level EncryptionRate LimitingCSRFCSPAudit Logging

Contact

Let's build something.

Whether you have a role, a project, or just want to connect — my inbox is open. I respond within 24 hours.

Message on WhatsAppConnect on LinkedInEmail me
Nairobi, Kenya+254 724 329 386github.com/gevversonAvailable for work