VeilPiercer - AI pipeline insurance for teams who ship to production
What it does
VeilPiercer traces every step in your AI agent chain - what each node READ vs what it PRODUCED - and lets you diff any two sessions to see exactly where they diverged. Add 2 lines of Python. No cloud. No Docker. No SaaS account.
The problem I kept hitting
I was running multi-agent LangChain pipelines overnight and had no idea why Monday's run and Tuesday's run gave different outputs from identical inputs. Traditional logs show you that it failed. VeilPiercer shows you where it split - which node, which read, which decision changed.
Debugging by reading raw logs at 2am is how you miss the $4,200/day token burn until it shows up in your Stripe bill.
What I built it on
- Local-first SQLite - your traces never leave your machine
- HMAC-chained audit log - tamper-evident by default
- Offline-first: works with Ollama, LangChain, CrewAI, AutoGPT out of the box
- Privacy gate: PII redacted at write time, before it touches any logger
Numbers from this week
- 18 gene pool variants evolved autonomously overnight (up from 6)
- 46 accepted production outputs traced and stored
- ~35% false-accept rate after CRITIC gate hardening (was 58.8%)
- PSO optimizer running on RTX 4060 tuning 8 live hyperparameters
- 0 cloud dependencies - entire stack runs on localhost
Pricing
- Free forever - pip install veilpiercer, traces locally
- $197 one-time - lifetime hosted dashboard, alerts, source code
- $499/mo - Team tier (5 seats, shared session library, REST API, audit log export)
Who's using it
Solo ML engineers debugging non-deterministic pipelines. DevOps teams running CI/CD agent pipelines. AI consultants who need to show clients "here's exactly what your agent decided and why."
What's next
- QLoRA fine-tune on 500 accepted outputs to eliminate formatting drift by default
- Qdrant semantic search over 92K session embeds
- n8n webhook: auto-trace every session on payment confirmation
-> veil-piercer.com . pip install veilpiercer . happy to answer questions
The phrase "AI pipeline insurance" is interesting, but teams shipping to prod will want to know if that means evals, guardrails, rollback, monitoring, or incident response. In practice, the hard part is proving it cuts bad outputs and pager time, not just adding another layer. A concrete example with before-and-after metrics would make this much easier to trust.