AI agent and backend engineer. Eight years of backend work, plus a Master's in AI.
Most AI integration work fails in the same few places: unvalidated input going straight into the model, the model's output going straight into a database, nobody counting tokens until the bill arrives, and no record of what happened when it got something wrong. I'm a backend engineer first, so those are the parts I build for.
Most of what I've built is private: client systems under NDA, internal platforms, production backends that will never have a public URL. The repositories below are the ones I can show. Each is a working system with tests and CI, built around the parts that are genuinely hard rather than the parts that demo well.
The permission layer for the "unvalidated input" and "no record" failures. A gate between an AI agent and the systems it can change: deny-by-default policy, scoped capability grants that expire, confirmation on anything consequential, spend ceilings, and a hash-chained audit log you can verify. The demo is seven scenarios, each one an attack and the guardrail that stops it — including the one most "confirm before executing" implementations get wrong: if you re-ask the model for arguments after approval, you approved a description and executed something else.
Go, standard library only. A TypeScript port of the confirmation core hashes byte-for-byte identically.
The answer to "nobody counted the tokens," applied to RAG specifically. Measures retrieval before you try to fix it: hit rate, MRR, precision and context cost per chunking strategy, against one goldset. Prints the rank each question's first correct chunk landed at and lists what the misses retrieved instead — an average tells you something is wrong, the miss list tells you what.
The answer to "output straight into a database." Measures a brand's existing
page, asks a model to assign roles and write copy, then forces that answer
through a closed schema where it loses every argument it has with a measured
fact. verify runs two different models over one source and exits non-zero if
the kits differ.
A production LLM extraction pipeline with an eval harness attached, not a demo. WhatsApp group messages to validated structured JSON across English, Gulf Arabic and Arabizi. Closed schema, mock LLM and form API so it runs with no key, 15/15 exact field match on the goldset.
TCP ingest gateway for GPS and telemetry devices. Sliding-window replay deduplication, mid-stream frame resync, half-open socket detection, and an explicit backpressure policy instead of an unbounded queue. Go, zero dependencies. Real-time, high-volume device data is my day job, not something I read about — I work full time on a telematics platform ingesting this kind of traffic.
Fleet tracking REST API. JWT auth, PostgreSQL with migrations, Redis caching, Swagger docs, Docker Compose.
Ethereum transaction parser with address subscriptions and block querying.
toolgate and device-gateway run gofmt, vet, the full test suite and the race
detector in CI, and cross-compile for arm64 — the target they'd actually ship
to. ragcheck and brandkit run on Python 3.9 through 3.13, and their CI
asserts the behaviour rather than just the tests: ragcheck gates the build on
retrieval quality, brandkit builds twice with two different models and fails if
the tokens differ.
Where clients bring me in: LLM features that have to work in production (OpenAI and Claude APIs, tool calling, RAG, streaming, retries and fallback, cost control), backend APIs in Python, Node, and Go, and the unglamorous parts most job posts leave out — idempotency, retries, duplicate messages, reconciliation, offline sync.
Available for contract work: Upwork