01
Production emergency: containing a $40K/day agent cost explosion
- Client.
- Late-stage platform, production multi-agent system.
- Engagement.
- Incident response · priced after triage.
The problem. A multi-agent pipeline in production began spending at a rate projected to exceed $40K per day — a $1.2M/month run-rate if left uncorrected — with no per-session budget controls and no circuit breaking at the tool-call layer.
What we shipped
- Led triage and spend attribution first: per-agent, per-model, and per-tool cost mapping before touching the live system.
- Shipped circuit breakers at the tool-call layer with per-agent and per-tool thresholds.
- Enforced per-session spend caps, keyed to task type, at runtime.
- Deployed retrospective governance: deterministic policy validation on every tool call, so spend is blocked before execution rather than reviewed after the bill lands.
Result
- Projected $40K+/day burn contained within 72 hours of engagement start.
- Zero recurrence since.
- The same controls shipped into the client's standard deployment path.
AWS Bedrock · AgentCore · policy-as-code tool-call validation
Engage on this: Production emergency · LLM Cost & Routing Sprint · Agent Production Audit
02
Platform Design & Build: multi-tenant agent substrate
- Client.
- Six-plus fast-growing platforms, Series B to late-stage teams.
- Engagement.
- Fixed-scope projects, ~6 weeks each, delivered to production and handed to the client's team.
The problem. Platforms shipping agents to production were hitting the same walls: shared infrastructure without tenant isolation, no spend control, no memory layer, and release overhead that slowed every iteration.
What we shipped
- Built the multi-tenant substrate agent systems run on — routing, memory, evaluation, and spend control — across AWS Bedrock, AgentCore, and Azure.
- Built Go, Python, and TypeScript middleware plus Azure DevOps deployment templates for four distributed engineering teams.
Result
- Delivered to production at 250-450 concurrent agent instances, then handed to the client teams.
- Release-cycle overhead reduced by roughly 40%.
AWS Bedrock · AgentCore · Azure · Go · Python · TypeScript · Azure DevOps
Engage on this: Substrate Build · Fractional AI Architect
03
Reliability program: SLO frameworks for three platforms
- Client.
- Three client platforms running multi-agent systems in production.
- Engagement.
- Reliability advisory · two quarters.
The problem. Teams were operating production agents without agreed reliability targets — every incident was argued over opinion instead of data, and improvements could not be measured.
What we shipped
- Defined SLO frameworks on task completion rate, mean tool calls, hallucination escape rate, and end-to-end latency.
- Tied error budgets to each objective, with agent trace instrumentation feeding the dashboards.
Result
- SLO compliance improved from 71% to 96% over two quarters.
OpenTelemetry · Prometheus · Grafana · agent trace instrumentation
Engage on this: Reliability Foundation — SLOs and evals in CI · Agent Production Audit
04
Retrieval and governance at scale
- Client.
- Platform running heavy tool-use agents in production.
- Engagement.
- Audit and build · production hardening.
The problem. A retrieval and governance layer that had grown ad hoc: no deterministic validation on tool calls, no visibility into what agents were actually invoking, and no evidence of enforcement for security review.
What we shipped
- Audited the retrieval stack on Pinecone and OpenSearch and rebuilt the ingestion and query paths.
- Enforced deterministic policy validation on every tool call, with zero known bypasses across deployments.
Result
- 1.8M+ tool invocations running under policy validation.
Pinecone · OpenSearch · policy-as-code tool-call validation
Engage on this: Agent Production Audit · Agent Governance & Insurability Package · AI Compliance Package — EU Article 50 plus US module
05
LLM cost reduction program
- Client.
- Portfolio of client platforms across the advisory engagements.
- Engagement.
- Cost optimization program · FinOps.
The problem. The same integration mistakes repeated across client codebases: context re-injected without caching, token-heavy integrations, and infrastructure patterns that multiplied spend.
What we shipped
- Refactored Anthropic SDK, n8n, and Copilot Studio integrations around token optimization, prompt caching, and low-infrastructure patterns.
Result
- $200K/yr aggregate reduction in cloud and LLM spend across the portfolio.
Anthropic SDK · n8n · Copilot Studio · AWS Bedrock
Engage on this: LLM Cost & Routing Sprint · Agent Production Audit
06
Data and ML platform standardization
- Client.
- Two client platforms with classical ML workloads.
- Engagement.
- Data platform standardization.
The problem. Data lakes and model registries that had diverged across platforms, making model lineage and dataset governance inconsistent.
What we shipped
- Standardized data lakes and model registries across both platforms, integrating Databricks and Domino Data Lab with existing workloads in scikit-learn, PyTorch, and TensorFlow.
Result
- One data and model lifecycle across both platforms.
Databricks · Domino Data Lab · scikit-learn · PyTorch · TensorFlow
Engage on this: Substrate Build · Fractional AI Architect
07
Real-time prediction-market trading system
- Client.
- Proprietary high-frequency trading firm · remote.
- Engagement.
- Systems engineering · Jan. 2026 - Jul. 2026.
The problem. A production trading system for prediction markets pulling live data from centralized exchanges and on-chain sources, with hard risk controls and sub-10ms peak latency requirements, trading real capital.
What we shipped
- Designed and ran event-driven Python and Rust pipelines holding sub-10ms latency at peak load.
- Designed the live risk-control layer: position limits, daily loss limits, and slippage checks enforced on every order.
- Built statistical models of market order flow to detect volatility clustering, and used them to size trades and set execution pace under fast-moving conditions.
- Implemented Kelly-criterion position sizing that scales exposure to model confidence.
- Maintained the research pipeline: historical backfill, detector-faithful backtesting, walk-forward validation, and automated signal outcome labeling.
Result
- Execution slippage cut by 14%.
- Zero limit breaches while trading real capital.
Python · Rust · centralized exchange APIs · on-chain data feeds
Engage on this: Reliability Foundation — SLOs and evals in CI · Agent Production Audit
08
Iguazio AI: inference platform engineering at scale
- Client.
- Iguazio AI (employer engagement · public on this site) · remote contractor.
- Engagement.
- Jun. 2021 - Mar. 2023.
The problem. Enterprise NLP and computer-vision inference workloads on a monolithic FastAPI service: 18-minute cold starts, a 900ms+ p99 on core analytics queries, and no canary rollout path.
What we shipped
- Deployed and operated a multi-GPU Triton Inference Server cluster with dynamic batching and model quantization, with inference SLOs and Prometheus/Grafana alerting runbooks for OOM failures.
- Decomposed the monolith into event-driven microservices on Docker, PostgreSQL, and Kafka.
- Owned schema and indexing strategy for behavioral analytics datasets across three product teams; redesigned indexes and materialized views.
- Engineered GitOps CI/CD with GitHub Actions and ArgoCD, automated health checks, and canary rollout policies.
Result
- Cold-start deployment time from 18 minutes to under 4.
- p99 query latency from over 900ms to 120ms.
- 99.9% uptime sustained across production containerized environments.
Triton Inference Server · FastAPI · Docker · PostgreSQL · Kafka · Prometheus · Grafana · GitHub Actions · ArgoCD
Engage on this: Substrate Build · Reliability Foundation — SLOs and evals in CI