- emailRe: re: leak under kitchen sink AGAIN
- formRent receipt for March, unit 12
- emailno heat since tuesday, baby at home
- emailQuestion about the parking permit
- formMaintenance request, 4B, front door lock
- emailFwd: noise after 11pm, 3rd floor
Agents that fit the way you work.
Agent systems forintake triage, grant reporting, invoice reconciliation, case notes, research briefs
Joinery is a one-person practice. I design, build and run AI agent systems for real workflows, each one cut to fit the tools your team already uses. Nothing ships without tests and a person signing off where it matters.
- sheet
- 01 / 09
- drawing
- intake to output
- scale
- 1 : 1
- drawn
- T. S. Bari
joinery, n.The craft of cutting parts to fit each other so they hold without nails.
Five jobs, drawn out step by step
These are example workflows, not client case studies. Each one shows the shape of what gets built: the inputs, the agent steps, where a person signs off, how the run is checked, and what comes out the other end.
Intake triage for a housing provider
A shared inbox of emails and web forms gets read, sorted and routed to the right team. Anything urgent waits for a person.
In this example: 43 messages sorted into four queues. The 3 urgent repairs wait for the intake lead before anyone is sent out.
Example workflow. Thirty-seven emails and six web forms arrive in a shared inbox. The agent reads each one, classifies it as a repair, tenancy, rent or general request, flags three urgent repairs, and routes everything into four queues. The three urgent items are held until the intake lead approves them. An inspection step checks routing accuracy, that every urgent item was caught, and that no personal data ended up in the logs. All three checks pass, and 43 tickets land in the existing helpdesk.
trace
0/11
- drawing
- wf-01 · intake triage
- status
- example workflow
- stations
- 5, one sign-off
- run
- waiting
Four ways to work together
Most engagements start with an audit. Some teams arrive with a build already scoped. Either way, the tests, the approval steps and the hand-over documents are part of the job.
s-01
Workflow audit
I sit down with your team, walk through how the work actually gets done, and mark the points where an agent could speed it up without lowering the bar.
- A time map
- where the hours go in one or two workflows
- A shortlist
- ranked by time saved, effort and risk
- A leave-alone list
- the work that should stay with people
s-02
Agent build
I design and ship an agent system on top of the tools your team already uses. No new platform to learn unless one is clearly worth it.
- shared inboxes
- spreadsheets
- CRMs
- document drives
- helpdesks
- accounting software
s-03
Evals and guardrails
The part that keeps an agent predictable after the demo. Every system ships with all three.
| Test suites | built from your own past cases |
|---|---|
| Approval gates | a person signs off where it matters |
| Monitoring | an alert when quality slips |
s-04
Operate and improve
Work changes, and the system has to keep up. After launch I stay on: the prompts and checks get tuned as your cases shift, problems get looked at the same week, and you get a short report each month on what the system did and where it struggled.
- weeklyreview flagged runs
- monthlyplain-language report
- as neededfixes and new cases
Four stages, shown as one dovetail joint being made. Measure: the two boards are drawn and dimensioned. Mark: the tail and pin lines are marked on the ends. Cut: the waste is removed. Fit: the boards slide together and seat.
Measure twice, automate once
A good joint is mostly measuring and marking. The cutting is quick once you know where the lines go. Every engagement follows the same four steps, and every system leaves with its tests, its checkpoints and its documentation.
step 1 of 4
Measure
Sit down with the people who do the work and time it. Where do the hours go, and which of them need a person's judgement?
step 2 of 4
Mark
Mark where an agent fits and where it doesn't. Agree what good looks like and how it will be tested before anything is built.
step 3 of 4
Cut
Build on the tools you already use, in small pieces. Each piece is tested against real cases from your own work.
step 4 of 4
Fit
Fit it into the day-to-day, with approval steps where they matter. It ships with its tests, its monitoring and documentation your team owns.
- Evals
- a test suite built from your cases, run on every change
- Human checkpoints
- named people approve the decisions that matter
- Documentation you own
- how it works, how to change it, what to watch
For mission-driven teams
Nonprofits, housing providers, social enterprises and public sector teams carry a reporting and admin load that would stretch a team twice their size. The aim is simple: give those hours back to the work you exist to do.
Reporting in the funder's language
A background in impact measurement means the frameworks are familiar ground, GRI, IRIS+ and SASB among them. Outcomes get mapped with care, and every figure points back to where it came from.
Care with the people in your records
When the records are about your clients and neighbours, privacy, consent and accountability come first. Personal data stays where it belongs, and a person signs off on anything that affects someone's service.
Sized for a lean team
Small, contained builds that one staff member can own after hand-over. No new platform you'd have to hire for.
Principles
- Measure twice, automate once
- Humans approve what matters
- Every agent gets tested
- You own the system
- No black boxes
- based in
- Toronto, Canada
- works with
- teams anywhere
- degree
- Master of Development Practice
Hi, I'm Tamjid
I build agent systems and then stay to look after them. Most of my recent work has been running Claude-based agents in production, which taught me that the build is the easy half. The harder half is the evaluation framework that tells you whether the agent is right, and the guardrails that stop it when it isn't.
Before that I worked in social impact: ESG consulting, community development and impact measurement. It's why I care about the unglamorous questions. Who signs off? Whose data is this? Can every number in the report be traced back to something that actually happened?
Joinery is how I put the two together. I sit down with your team, find the places where an agent would genuinely speed the work up, and build it so the quality holds.
Built an agentic workflow for the QA team at a financial services provider that cut regression testing from two days to twenty minutes.
| part | step | before | now |
|---|---|---|---|
| A | Pick the cases. Read the release notes and changed tickets, then select the regression cases they touch. | 1 h | 1 min |
| B | Set the bench. Reset the test environment and load synthetic accounts, balances and statements. | 2 h | 3 min |
| C | Run the suite. Agents work through the selected UI and API cases in parallel, capturing evidence for each. | 10 h | 11 min |
| D | Sort the failures. Re-run flaky checks, separate real defects from noise, and recheck totals against expected values. | 2 h | 3 min |
| E | Write it up. Draft defect tickets with steps to reproduce, logs and screenshots attached. | 1 h | 2 min |
| F | Sign off. The QA lead reviews the findings and decides whether the release goes out. | stays with a person | |
| total, one pass | 16 h | 20 min | |
Questions people ask first
Something else on your mind? hello@joinery.fyi
It starts with a conversation. We sit down, you walk me through the work, and together we find the points where automation would speed things up without lowering the quality. From there it is usually an audit, then a build in small pieces, then a stretch of running it together before your team takes it on. You see the scope and the price before anything starts.