01 / SERVICERAG that holds up under eval
Retrieval-augmented generation with the eval harness built in — an index is the right call on layout-heavy, citation-bound, or latency-tight work; the diagnostic makes that call before the build does, not an assumption baked in on day one. We pick the chunker, the embedding model, and the retriever — contextual chunking and late-interaction retrieval where they earn their place — and benchmark every change against a golden set you'll keep using after we leave. The goal is fewer unsupported answers, less manual review, and degradation caught before customers see it. The same harness ships on its own, for a model already in production with no way to know when it degrades.
- ── Hybrid + late-interaction retrieval, agentic re-query
- ── Contextual & late chunking per document class
- ── ColPali visual retrieval for layout-heavy docs
- ── Faithfulness + groundedness evals, CI-gated
pgvector · bge-m3 · cohere-rerank →
02 / SERVICEAgentic harnesses
Multi-step agents with tool use, tiered memory, and budgets that don't blow up production costs. Built on MCP, gated by evals, traced end-to-end — and that observability and eval layer retrofits onto an agent you already run. The point is bounded spend, bounded blast radius, and failures your team can inspect instead of replaying from logs after the fact. When the agent has to spend money, the payment-rails service ships spend policy outside the model — ceilings, treasury isolation, and circuit breakers the agent cannot reason around.
- ── Tool orchestration over MCP servers
- ── Budgets, replanning, structured failure contracts
- ── Tiered memory — working, episodic, semantic
- ── Trace-level observability + per-step evals
LangGraph · MCP · Temporal · Langfuse →
03 / SERVICEVoice agents
Streaming voice systems on phone trees, kiosks, and apps, architected to a sub-300ms p95 budget. STT → LLM → TTS with model-based barge-in, drift detection, and PII-safe transcripts. The system is built to cut handle time without hiding latency, failed turns, or policy-risk calls from the operators who own the queue. Every component on a millisecond budget.
- ── Sub-300ms p95 latency budget
- ── Model-based barge-in + back-channeling
- ── Call recording + drift detection
- ── PII-safe transcripts
LiveKit · Deepgram · ElevenLabs →
04 / SERVICEPost-training & grounding
Fine-tuning and post-training on your data and your task — for when prompting and retrieval have hit their ceiling. SFT, preference tuning, and eval-gated checkpoint selection. We make the gain reproducible, not mystical: the training recipe is handed off, so you can rerun it, audit it, and keep improving the model after we leave.
- ── SFT and preference tuning on your data
- ── Eval-gated checkpoint selection
- ── Distillation for latency and cost
- ── A reproducible training recipe at handoff
TRL · vLLM · Modal →
05 / SERVICEAI-agent security & audits
Security audits for agents that hold wallets and sign transactions. We red-team the prompt-injection-to-transaction attack surface that smart-contract auditors don't cover — because the contract can be fine while the agent is still the hole. The work gives finance, security, and engineering a shared view of what the agent can spend, sign, and refuse.
- ── Prompt-injection → transaction red-teaming
- ── Spend-limit and refusal-boundary review
- ── Signing-key isolation + MCP allowlist audit
- ── ERC-8004 identity hygiene
Foundry · ERC-8004 · custom injection suites →
06 / SERVICEAgent treasury & payment rails
The payment protocols for agents — x402, ERC-8004, AP2 — shipped the rails to hold and spend money. The controls that keep a prompt-injected agent from draining a wallet did not. We build that layer: tiered spend ceilings enforced outside the model, treasury isolation with no auto-top-up, drawdown circuit breakers, and immutable receipts your ops team can actually investigate. The agent gets frictionless per-call payments; finance and security get a bounded blast radius they can sign off on.
- ── Per-call, per-counterparty, and per-day ceilings enforced outside the model
- ── Treasury isolation: hot-wallet float only; no agent-initiated top-up
- ── Drawdown circuit breakers and immutable payment receipts
- ── x402 / ERC-8004 / AP2 integration wired into the agent harness
x402 · ERC-8004 · AP2 · Foundry →
