Work
Our own products, open-source tools, and client case studies — some anonymized under NDA. What we show is the engineering scope delivered, not promised business numbers.
No projects match this combination.
Company Website for a Design-Build Contractor
Nine routes live with an email-first lead pipeline, WhatsApp fallback, and CI-enforced media budgets
Agentic Engineering: Risk-Routed Spec-Driven Dev
A spec-driven pipeline published as a plugin: risk-based T0–T4 routing and generated adapters for eight other agent harnesses
Jelma: An Indonesia-First AI Website Builder
A 25-feature PRD decomposed into 10 epics and 41 stories; the app foundation and build pipeline run, with the product in active development
Pandu: Self-Hostable RAG, Measured End to End
Hybrid retrieval measured at recall@5 0.967 on the NIST golden set, with CI gates that fail on any regression — including one gate deliberately left red
Greenlight Console: Every Write Goes Through a Human
Every write operation pauses on an in-chat approval card and runs only via a single-use, argument-bound token minted by a human click
Hybrid RAG Reference — Benchmarkable Hybrid Retrieval
Nine-configuration benchmark over 120 labelled queries; re-measured against real Gemini embeddings
LLM Cost Gateway — A Spend Gate for LLM Providers
222 offline tests; verified live end to end against DeepSeek on 2026-08-16
LLM Eval Harness — Prompt-Regression Diffs for CI
179 offline tests; verified live against DeepSeek, gemini-2.5-flash-lite, and an Anthropic-compatible endpoint
LLM Flightdeck — Mission Control for LLM Operations
Four KPI pages; every figure recomputed from real run artifacts of three upstream repos
MCP Guarded Tools — Large Tool Surface Governance
43 tools behind 4 meta-tools; tool context measured 83.3% smaller live; 62 offline tests
Notula: A Measured Two-Pass Meeting Notes Pipeline
A transcript as system of record makes bilingual EN/ID summaries re-runnable without touching the audio again, with per-stage cost metering in the API
Prompt Injection Bench — Measuring Defenses
127 attacks × 9 presets × 5 repetitions (5,715 mock attempts); plus a live measurement against deepseek-v4-flash
Roundtable Agents: Multi-Agent Research on a Graph
A complete research pipeline with restart-proof checkpoints and a measured solo-vs-multi comparison published either way, including its negative result
MBS Food Cost: Ops System for Multi-Outlet F&B
Full parity with the legacy system, held by a regression invariant: true food cost ≈45.3% against the ~66% the business believed
CloseLoom: AI-Native CRM for Indonesian B2B
Multi-tenant foundation in place: continuously tested RLS isolation, two foundation stories done, the rest a documented backlog
Document Processing Agent for Logistics Operations
Three document flows running automatically with layered validation and an operator escalation path
InvestorID: AI Investment Intelligence for the IDX
A daily 900+ IDX stock pipeline running serverless on Cloud Run + Supabase, at ~$0.10/month per the README's cost table
Data Platform and Dashboards for a Finance Provider
Four source systems consolidated into one warehouse with automated quality checks on every load
Prompt Enhancer: A Chrome Extension for AI Prompts
A shipped extension working across four AI platforms, with secure API key storage and right-click context-menu integration
Some client case studies are anonymized under confidentiality agreements. Client names and business figures are not published; our own products and open-source tools are shown as they are.