foundry · zaakir.io github.com/zaakirio

Foundry

One portfolio, eight measured AI-engineering systems.

Foundry is the connective map for the product family at *.zaakir.io. Flux binds the fleet together with trace, cost, registry, and eval health. Quench and Anvil are agent nodes. Ingot accounts for tokens against delivered work. Sluice routes model calls, Assay proves extraction with per-field evals, and the Crucible/Bellows pair keeps the local-inference specialist story contained and measured.

Start with Flux Inspect the lineup
Foundry product map with Flux at the center connecting the eight products
fleet map · live flux binds one trace across 2 languages drag to orbit · click a node to open

The lineup

The portfolio is arranged as a fleet, not a pile of demos. Each project owns one screened AI-engineering signal and exposes enough evidence for a skeptical reader to audit it.

flux
Vendor-neutral control plane for heterogeneous agent fleets.
Proves one stitched OpenTelemetry trace across Quench (TS) and Anvil (Python), with per-agent cost and eval-health roster.
quench
CI/deployment incident-triage agent with human-gated writes.
14 trajectory evals, 78 checks, 12 real recorded GitHub Actions failures, dry-run by default.
anvil
Support agent over real FastAPI documentation with retrieval and answer evals.
Hybrid retrieval, rerank, grounded citations, HITL issue tools, recall@5 0.567 on real questions.
ingot
Token-to-deliverable analytics for AI coding agents.
Joins agent usage telemetry with git history: spend, cache economics, cost per shipped commit.
sluice
Cost-aware OpenAI-compatible gateway and router.
Routing policy, fallback chains, retries, circuit breaker, SQLite cost ledger, OTel and Langfuse traces.
assay
Enterprise document extraction with per-field evals.
Real SROIE receipts, controlled synthetic stress set, constrained decoding, validation, confidence routing.
crucible
Forensic eval harness for self-hosted models.
Capability, refusal behavior, tool calling, RAG, base-vs-abliterated deltas, judge grading, CI gates.
bellows
MCP server for operating a local llama.cpp fleet.
Typed tool calls to scan GGUF models, supervise llama-server, smoke-test, and query Crucible eval history.

How to read it

Start with the system-level pane, then follow the evidence down into the agents, evals, cost accounting, and model-internals niche.

  1. 01
    Flux

    The keystone. It shows the fleet as one operational system instead of separate repos.

  2. 02
    Quench and Anvil

    The live agent nodes. One handles incident triage; the other grounds answers in real docs.

  3. 03
    Ingot and Sluice

    The economics layer: what tokens bought, and how requests route under cost and reliability constraints.

  4. 04
    Assay and Crucible

    The eval story: per-field document extraction and model-level behavior measurement.

  5. 05
    Bellows

    The local-fleet MCP tool, deliberately kept as a specialist support story rather than the lead.

foundry / interview map
evals engineering        anvil, assay, crucible
agent orchestration      quench, anvil, flux
MCP artifact             bellows, anvil surface
RAG metrics              anvil
cost accounting          ingot, sluice, flux
observability            flux, sluice, quench
reliability boundaries   quench HITL, sluice breaker
local model specialist   crucible, bellows

Foundry is an index and narrative layer, not a ninth product claim. The product repos remain the source of truth.

Limits, stated plainly

  • This is a portfolio and product family, not a hosted SaaS suite or a company homepage.
  • Several metrics are measured offline or on one machine. The linked product repos explain the exact setup and limits.
  • The product sites summarize evidence; the READMEs, fixtures, eval outputs, and code remain authoritative.
  • Crucible and Bellows are intentionally the local-inference niche. The portfolio lead remains cloud-API production systems because that is where the 2026 market screens hardest.