I build and run production AI — agents, LLMOps, RAG.
Six years shipping ML and LLM systems end to end, plus the backend and
infrastructure underneath them. Currently founding engineer at MayaTravel
and building AcruxCore.
A production full-stack AI SaaS I built alone, shipped and open-sourced. The thesis: assume the cheapest model you can stand, then engineer hard around its weak spots.
Made cheap LLMs reliable — batched segments, enforced one-to-one output counts with retries, and always restored the original timestamps so a hallucination can never shift the timeline
Pluggable ASR providers chosen per job from config, with a health check that fails over automatically so idle GPU capacity costs nothing
The stateless Python AI worker never touches the database; the Node backend owns all state, identity and quotas
Three years as founding engineer, owning the AI system the product runs on — the agent graph, retrieval, prompts, per-client behaviour, and the tracking that makes all of it debuggable in production.
Wrote ~25,000 of the ~38,000 lines of the Python AI service over three years, then led porting the whole thing into the Node/NestJS backend — two languages were creating friction and slowing the team down
Moved prompts into the database with versioning and a safe fallback, so prompts and models change in production without a deploy
Built per-client behaviour for 35 enterprise clients out of one shared codebase — intent routing, prompt variants, retrieval grids, feature flags
Built the retrieval layer — hybrid metadata search, filter tracking, quick search and fallback chains — plus tool-level tracking so an agent's decisions can be debugged after the fact
Automated client feedback handling: feedback from client users is triaged and actioned through Claude Code skills I built for it
Grew it from the first B2B client to 30+ enterprise deployments, with client retention holding throughout
Soffos.ai
NLP Consultant
2022 — 2023
USA — remote
Built the NLP services behind the Soffos SDK.
GPT-3 prompt engineering plus FastAPI microservices for classification and ambiguity detection, shipped into the public SDK
Document parser extracting structured content from complex layouts
Founding engineer at a telemedicine startup during COVID. Owned both the model and the cloud infrastructure it ran on.
Remote vital signs — heart rate, respiratory rate, oxygen saturation — from a phone camera video feed in PyTorch, to ±5 MAE on heart rate
Ran the data collection myself: recorded and labelled a dataset from hundreds of participants, because no suitable one existed
Built the AWS stack myself (Lambda, S3, EC2, SageMaker) and integrated the model into the live doctor-consultation workflow
Freelance track record
2019 — 2023
Years of client work running alongside the full-time roles — computer
vision, NLP, speech and data pipelines, delivered directly for clients who
rated it and came back.
