 |
Wildchild
Forward Deployed Engineer · Sep 2026 - Present
Wildchild AI OS: Humano's multi-tenant agent and application platform. Meridian is the imaginary reference client used for evals and demos.- Led the cutover of the runtime and workflows from Python to native .NET 10 / ASP.NET Core, removing 37K+ lines of Python. Shipped a Python-free image, an Aspire one-click local platform, and native production on GKE with digest-based release, backup, rollback, and Cloud Monitoring.
- Built the agent surface clients use: a version-pinned skills library (max 12, fails closed), custom connectors (BigQuery, Composio, HTTPS JSON, remote MCP over streamable HTTP, SSRF-safe with row, time, and size caps), builder chat across devices, shared project rooms, and per-prospect workspaces. Several are in review as of Nov 2026.
- Evals gate what ships. Two frozen Langfuse datasets on synthetic Meridian data: context-graph Q&A went from 2/10 to 9/10 and invoice routing from 19/40 to 38/40, with gates on exception recall and log precision. Online LLM-judge scores, self-hosted Langfuse, OTel into Cloud Trace.
- Each client gets an immutable OCI image, never the source repo. Deployable into client clouds behind provider seams; Keycloak OIDC and Google IAP for tenancy.
|
 |
Humano
Engineering Lead / Forward Deployed Engineer · Sep 2024 - Present
Product owner and technical lead on client projects. Manage two QA engineers and two FDEs; run hiring. Joined for the LiveTwin MVP; now lead across Humano products and client work.
Internal: shared Jenkins CI/CD (merge to Cloud Run in 45-70s), Foundry (ticket to coding agent to green PR), Fyndry Slack agents with secrets kept out of sandboxes. |
 |
Equisoft RFX
Humano client · InsurTech, Canada data residency
- Challenged a SharePoint-only scope in discovery; the client reframed the win as fewer wrong answers, faster. Agreed weekly feedback (~50 answers/week for ~3 months) and mockups before UI.
- Shipped Azure Container Apps (Bicep/azd), Entra SSO, IDOR/SSRF-gated APIs, bulk Q&A upload, DOCX extraction of 121 requirements (62 tests, Azure prod).
- Answers draw on 1,200+ approved Q&As: exact match, then near match, then 70/30 hybrid RAG, then a human, with a claim verifier and Langfuse. Hotfixed an import that read 21 of ~409 questions.
|
 |
Bumi
Humano client · Malaysian consumer lending, in production since July 2026
- Constraints: BNM audit logging, maker-checker, 7-year WORM retention, 60% DSR ceiling, a partner bank portal with no API.
- Shipped Innov8tif eKYC, headless portal sync, multi-bank routing, GPT-4o-mini document vision and a decline classifier (humans on doubtful cases, never auto-reject), WhatsApp/SMS OTP, Pulumi GCP, Jenkins CI, and a 13-finding pentest remediation.
- About 470 API and Playwright specs gate each release; I own CI, staging, QA, and client retest.
|
 |
LiveTwin.ai
Humano product · app.livetwin.ai
- Production owner (Vercel, Infisical).
- 25-tool MCP server, Tavus real-time avatars, Neo4j and Qdrant memory, Salesforce/HubSpot via Composio, 96.2% statement coverage.
- Shipped a client product-expert avatar.
|
