AI Agents & Workflow Automation
I turn flaky AI agents into ones you can ship, and I automate manual workflows with LLMs. The work is the unglamorous part, namely tool calling that validates, retries, evals, and a human in the loop where mistakes are expensive.
Sound familiar?
- ✕The agent works in the demo and breaks on real inputs
- ✕Tool and function calls fail silently and the agent keeps going
- ✕You can't tell why it failed because nothing is logged or traced
- ✕A task takes 10 manual steps and you want an LLM to do most of it
- ✕You're scared to let it act on anything because mistakes are costly
- ✕Every prompt tweak fixes one case and quietly breaks two others
What you get
- ✓An agent with a measured reliability number you're comfortable shipping
- ✓Tool calls that validate inputs and outputs and fail loudly
- ✓Retries, fallbacks, and a human checkpoint where a mistake is expensive
- ✓A manual workflow automated end to end, with a person in the loop
- ✓Tracing and evals so you can see what the agent did and why
A 70% agent is a great demo and a poor product
Getting an LLM to call a tool once in a notebook is easy. Getting an agent to do real work reliably, handle bad inputs, recover from a failed call, and stop before it does something expensive, is the hard part. That gap is mostly engineering, not prompting. I spend most days running Kubernetes and shipping AI infra, and I’ve watched plenty of impressive agent demos fall apart the moment real users fed them real inputs. The fix is validation, retries, evals, and a human in the loop where it counts.
What I help with
- Reliable agents. Tool and function calling with schema validation, retries, fallbacks, and a clear stop condition when the agent is unsure.
- Workflow automation. Manual, repetitive tasks (triage, routing, data extraction, drafting) automated with an LLM doing the reading and writing.
- Human in the loop. Approval checkpoints and review queues for any action that’s costly or hard to undo.
- Evals and tracing. A fixed task set to measure reliability, plus traces so you can see every step the agent took and why.
- Guardrails. Narrow tool permissions, spend and rate limits, and audit logs so an autonomous agent can’t run off the rails.
Why me for this
AI infrastructure is my day job at TrueFoundry. I work on the serving and gateway layers that agents depend on, plus the observability that tells you what an agent actually did in production. You get someone who builds the agent and understands the system underneath it.
How an engagement works
- Scope. The task to automate, the inputs it sees in the wild, and the reliability bar that makes it worth shipping.
- Build the loop. Tool calling with validation, retries, and fallbacks, kept as simple as the task allows.
- Measure and harden. Run it against a real task set, add the human checkpoints and guardrails, and push the reliability number up.
- Hand off. Traces, evals, and runbooks so your team can extend and operate it without me.
Frequently asked questions
Why is my AI agent unreliable, and how do I fix it?+
A 70% agent usually fails on the connective tissue, not the model. Tool calls return junk and nothing validates it, one bad step cascades into a worse one, and there's no retry or fallback when a call fails. I add schema validation on every tool input and output, retries with sensible backoff, and a hard stop or human handoff when confidence is low. Then I measure it on a fixed task set so "more reliable" is a number, not a feeling. That path usually moves an agent from the high 70s into the 90s.
What kinds of workflows are worth automating with an LLM?+
Anything that is high volume, follows loose rules, and involves reading or writing natural language. Triage and routing, drafting replies, extracting structured data from messy documents, summarizing, and first-pass classification are strong fits. Tasks that need perfect accuracy every time or carry legal or financial risk should keep a human in the loop. I help you pick the workflows where the economics actually work.
Should I use a framework like LangChain, or build it directly?+
It depends on how complex the agent is. For a tool-calling loop with a handful of tools, calling the model API directly is often simpler to debug and keeps you in control. Frameworks help once you have many tools, multi-step planning, or shared memory, but they add abstraction you have to learn and trace through. I pick based on your actual complexity rather than the trend of the month.
How do you keep an autonomous agent from doing something harmful?+
Limit the blast radius. Give the agent the narrowest set of tools it needs, validate every action against a schema before it runs, and require human approval for anything destructive or costly. Add spend and rate limits, log every action for audit, and run the agent against an eval set of adversarial cases before launch. Autonomy is a dial, not a switch, and I set it where the risk is acceptable.
related work
Where I’ve done this
Built a customer-support AI agent that deflected ~60% of tickets
A support team was drowning in repetitive tickets that the help docs already answered. An AI agent over those docs and past tickets, with order-lookup tools and a clean human handoff, took ~60% of volume off the queue.
~20 hours a week of manual ops work eliminatedAutomated a manual ops workflow and saved ~20 hours a week
An ops team hand-sorted every incoming request, reading, categorizing, and routing each one by hand. An LLM pipeline that triages and routes with a human review step gave back ~20 hours a week.