01
Sit with the work first.
Before any code, we watch the actual workflow. Where does it slow down, what does a bad day look like, what would a sensible human do?
All services
Primary practice
AI Automation & Agents
Quiet software that does the boring half of the job. Built to hold up under real load, not just demo well on a laptop.
What we're replacing
A typical Monday at a 40-person ops team
Before
- 08:47Opens 7 tabs to start the day12 min
- 09:15Copy-pastes 80 vendor rows into a sheet45 min
- 10:30Emails 6 vendors the same question20 min
- 11:00Fixes one typo across 12 places25 min
time burned
3h 12m
After we ship
- 08:47Reviews a Teams summary3 min
- 08:50Done. Coffee.
time touched
3 minutes
Across a team of eight, that's ~25 hours back every week.
The pain
Smart people, spending half their week on work a machine could do.
Copy-pasting between four tools to assemble one report.
Reading the same six emails every Monday before anything starts.
Fixing the same typo in twelve places, hoping you caught them all.
High stakes if something gets missed — and something always gets missed.
Approach
02
Ship the smallest useful thing.
Build for the predictable cases first. When the system isn't sure, it pauses and flags it, with full context attached. Costs, retries, and logs visible from day one.
03
Run the new beside the old.
If a model's involved, the new one runs alongside the incumbent. We only switch when the new one actually wins on the measures that matter. No blind swaps.
04
Design for bad days.
Feeds drop records, services blip, things go slow. A summary that takes eight seconds is a summary nobody reads. We design for those moments, not around them.
Initiatives
Concrete things we've built or are building, against this kind of pain.
01 / Production
Anomaly detection at scale
A Z-score-based price anomaly system running over 1M+ Amazon FBA records, surfacing outliers automatically rather than via manual scanning.
FBA Research Platform02 / Pattern
Document & data processing pipelines
Pattern: messy unstructured input (PDFs, spreadsheets, emails) into a queryable structured store with an LLM doing the parsing and an audit trail behind it.
03 / Pattern
Internal copilots & knowledge agents
RAG over a client's internal documents so their team can ask questions in plain language instead of digging through files.
Roadmap
How an engagement typically unfolds. Not rigid, but the shape repeats.
Phase 01
Map the workflow. Find what's actually slow and where mistakes happen. Decide what to skip.
Built with
- OpenAI
- Anthropic
- LangChain
- Vector DBs
- Python
- TS