01.52sguardrailamount over approval limit -> human review
01.53shandoffap-review queue · summary attached
6 steps · 1.53s · $0.004 · evals 14/14 passing
Representative trace
Engineering alongside teams at
BlackBuck
FourKites
StockOne
KPMG
Pine Labs
YES BANK
TiVo
Xperi
Cogoport
Miebach
Pando
NextWealth
worxogo
01 / The shift
The demo worked. Production is a different problem.
Most enterprise AI stalls between a convincing prototype and a system people can depend on. The model is rarely what is missing. What is missing is the engineering around it: clean ground truth, tool access, fallbacks, evals and someone who owns the outcome.
Autonomous and human-in-the-loop agents that use your tools, call your APIs and reason through multi-step work, with deterministic fallbacks when the model is unsure.
TOOL USE · API ORCHESTRATION · APPROVAL GATES · FALLBACK PATHS
Supply chain, logistics, finance and enterprise ops run on spreadsheets, inboxes and brittle scripts. We replace them with reasoning-driven pipelines that handle the exceptions rules cannot.
Static dashboards answer last quarter’s questions. We build interfaces that turn natural language into queries, executable code and operational views composed on the fly.
›Which lanes missed SLA last week, by carrier?
SELECT carrier, count(DISTINCT lane) AS lanes, count(*) AS missed FROM shipments WHERE sla_missed AND ship_date >= current_date - 7 GROUP BY 1 ORDER BY missed DESC
High-concurrency crawling, document intelligence and hybrid retrieval, so your models reason over clean, current facts. Headrun has built web-data infrastructure since 2010.
Continuous eval harnesses, latency and cost routing, security guardrails and observability. This is what separates a demo from a system you can put your name on.
P500.8sP952.1sCOST / TASK$0.004EVAL PASS98%
EVAL HARNESS · MODEL ROUTING · GUARDRAILS · OBSERVABILITY
Every system is built by engineers embedded in your team.
A forward-deployed engineer sits with your team, in your tools, on your problem. They diagnose the bottleneck, design the system and ship it. Then they stay until it holds.
One call, one page. We name the workflow, the metric and the first thing we would ship.
Gate 1 · Metric agreed
Phase 2
Embed
Engineers join your standups, your repo and your Slack. They map the workflow as it really runs, not as it is documented.
Gate 2 · Workflow mapped
Phase 3
Ship
A working system in production on a narrow slice of real work, with evals from the first commit.
Gate 3 · Live on one slice
Phase 4
Harden
We widen the scope, tighten the guardrails and hand your team the runbook. Or you keep us on.
Gate 4 · Runbook handed over
your-org/ops-platform · mainHHeadrunYYou
Hday 02 · docsmap invoice-exception workflow as it runs today
Hday 04 · featretrieval over contracts and rate cards
Yday 08 · evaladd 40 golden cases from last quarter's exceptions
Hday 11 · featexception agent, shadow mode
Yday 15 · fixfallback when the rate card is missing a lane
Hday 19 · deployproduction canary at 10% of invoices
Hday 24 · deployproduction, full volume, human review on high value
Representative engagement
Three ways to buy AI engineering
Consultancy
You get a strategy and a roadmap. Someone else has to build it.
Staffing firm
You get headcount. You still own the architecture, the management and the outcome.
Headrun
You get a running system, the evals that prove it and engineers who answer for it.
04 / Blueprints
The patterns we deploy, in the open.
These are architectures we put into production: the components, the failure modes and the numbers we tune for. Take them, or have us deploy them.
Idle
Trigger (ERP exception). Picks up the exception event from the ERP. Fails when: Events are duplicated or arrive out of order. Safeguard: One idempotency key per exception.
Retrieve (Contracts, rates). Pulls contract terms, rate cards and prior decisions for the case. Fails when: The relevant document is missing or stale. Safeguard: Freshness check. An empty result goes to human review.
Agent (Plan and act). Plans the resolution, calls tools and proposes an action with its reasoning. Fails when: A tool errors or the model is unsure. Safeguard: Step and time limits, then the deterministic fallback.
Guardrail (Policy check). Checks the agent’s proposed action against policy before anything is written. Fails when: The policy is stale, or an amount was misread upstream. Safeguard: Anything over the approval limit goes to a person. Every block is logged with its reason.
Write (Post to ERP). Posts the approved action to the ERP with an audit record. Fails when: The write partly fails. Safeguard: Transactional write, retried with the same key.
Fallback (Fixed rules). Takes over when the agent cannot finish: holds the case, attaches what was found and routes it to a person. Fails when: A rule no longer matches how the process runs. Safeguard: Rules are versioned and replayed against the golden set.
Human review (Approve or edit). A person approves, edits or rejects, with the agent’s summary attached. Fails when: The queue backs up. Safeguard: Alert on queue age. Decisions feed the golden set.
Evals and tracing (Logs every step, replays the golden set before each release). Logs every step and replays the golden set before each release. Fails when: The golden set drifts from real cases. Safeguard: Refresh it from reviewed cases on a schedule.
Tuned for · Auto-resolution rate, time to human decisionReference design
Headrun has built software and data systems since 2010, beginning with web-scale crawling and extraction. A client’s numbers appear here only with their sign-off.