AI/Application Engineer. I build AI systems that take real actions in production, and the controls that make that safe.
Six-plus years shipping backend and full-stack software. The last stretch has been agents, retrieval and evaluation, on top of a core banking platform and a multi-product API I still run.
Kigali, Nigeria and Dubai at different times. Remote worldwide, and I travel.
trackline watches whether an AI agent is still doing what you asked. An agent that drifts does not crash: it writes to .env, installs a package nobody named, edits code outside the request, and the build stays green. trackline runs beside Claude Code, Codex and Cursor, names what it saw with the evidence, and can stop the action before it happens. The same engine checks production agent traces, and a CI eval gate fails the build on regression. It belongs to no vendor, so one set of rules covers every agent a team uses. Go engine, open source, on npm (npm install -g trackline), with a phone app to approve or block what it stopped. → trackline.dev · why I am building it
It came out of Lira Intelligence, an AI support agent that answers from a company's own knowledge base and then does the thing: freeze the card, check the transaction, open the ticket. The interesting part is not the answering. It is letting a language model take privileged actions without that being reckless.
So it has:
- Production RAG on Qdrant. Hybrid search, vector plus keyword, ranked by source authority and filtered by knowledge-base segment. It degrades to keyword retrieval when the vector store is down instead of failing the turn.
- A seven-tier risk model on every tool, from
read_publictohuman_only, plus maker-checker approval: the agent parks a privileged action, a second person authorises it out of band, and the tool executes server-side and writes back into the conversation. An approver acts hours later, when the customer's socket is long gone. - An MCP gateway so a customer can plug in their own tool server. Off by default, KMS-backed credentials, SSRF protection through DNS resolution and private-IP blocking, per-tool rate limits, full config audit trail. → how and why I built it this way
- A realtime voice agent in Python on Pipecat with AWS Nova Sonic.
- An eval harness that fails the build on regression. → open sourced, and now the CI half of trackline
It runs on one Fastify service on AWS serving four products from 155 service modules and a 46-model PostgreSQL schema, isolated by table prefix and product-scoped JWT claims.
Separately I lead backend on a production core banking platform in .NET 8. Over the 30 days to 9 September 2026 it served 65,942 requests with two server errors, p95 284ms. That same governed agent now runs inside its API.
Client systems and commercial products do not go on GitHub. What I can do is open source the parts that carry no customer data, and write up the architecture.
| trackline | The alignment layer above. Claude Code, Codex, Cursor, any MCP client and production traces, from one engine. Its CI eval gate gives safety metrics an absolute floor of zero, so an agent that starts complying with prompt injection cannot pass on tolerance. |
| agentfile | Writes the context file an AI coding agent reads before it touches your code. No backend, no keys, no model call. |
| liracall | An AI voice agent that feels like a real phone call. Native call screen, live cloud backend. |
| cvault | Privacy-first VPN on WireGuard. Backend, desktop client, JavaScript SDK and web demo. |
| yerinsabraham.com | My site, and where the engineering write-ups live. |
The contribution graph counts private work, which is most of it.
AI RAG, agents, tool calling, MCP, evals, guardrails, prompt-injection defence, realtime voice · Anthropic Claude, OpenAI, AWS Bedrock, Nova Sonic, Pipecat, Qdrant
Backend TypeScript, Node.js, Fastify, Python, C#/.NET 8, Go · PostgreSQL, Prisma, DynamoDB, Redis, Kafka
Cloud AWS (EC2, ECS, RDS, S3, KMS, Secrets Manager, Bedrock), Docker, GitHub Actions, CI/CD
Frontend React, Next.js, TypeScript, Flutter
I am a medical doctor, and I am building Oystar, which carries a patient's case from a frontline clinic to the right specialist and brings the clinical answer back. Live in Rwanda. Patients who never arrive get flagged instead of lost, and the patient needs no phone and no app.
I run Creovine Academy, teaching people to work with AI. 120+ people so far.
I have been building since 2020, under the Creovine name since 2025. Before that I founded Metart Africa (2022), a web3 platform for African art, and spoke for it at Nigeria Fintech Week in Lagos that October. I also built Adna, a crypto payment gateway for Nigerian merchants. My work on the Riverly banking platform is under @Yerinsfluxus.
Outside all of it I draw, in pen and ink, at some scale.
Open to senior AI and backend engineering roles. Remote worldwide, EOR or contract.
Every claim on my CV, with where to check it: yerinsabraham.com/verify