Parham Davari

Oraclous is building a platform where people and AI agents work on the same team. It hasn’t launched yet. I thought small businesses without their own tech team were the ones who needed it most, so I led the backend for Apps. It lets a business owner take an AI-agent workflow and turn it into a simple app their employees can use.

A lot of my time goes into keeping agents reliable. Long runs kept timing out while they waited for a person to approve something, so I measured how slow the models really were and sized the time budgets from that. I also help with GraphRAG on Neo4j, where search mixes semantic and full-text matching and only ever looks at one organisation’s data.

My favourite bug so far: fastapi-loopguard, a tool I’d built on the side, caught something that stalled every auth request for up to 1.5 seconds on each login. I fixed it and added a test so it stays fixed.

Aawiz is a startup making an AI companion that looks after people’s wellbeing and helps teams keep track of their mood. I joined on a four-month part-time contract, designed and built the MVP backend, and worked out the features with the product manager and the designers as we went.

The companion itself was the fun part. It remembers what you talked about, replies live, and sends mood reminders, but never during your quiet hours. Under the hood it’s Django REST Framework with pgvector, Redis and Celery, running on Azure. You can try it at dev.pwa.aawiz.net.

Mach Milling makes dental prosthetics and runs dental clinics, with over 100 people in all. I was their first software hire.

When the internet was shut down, orders fell to a quarter of normal. We built a Django ordering platform and launched it in 10 days. Orders climbed back past where they’d been before, and it still takes about 150 a day.

After that I designed and built Rasa, which moves those orders through production. It’s a set of event-driven FastAPI services on Redis Streams, with the infrastructure in Terraform, and we piloted it in-house.

Some of my favourite work there wasn’t code. I crawled and analysed six years of data on 43,000 patients at Nikan, the group’s clinic, and what we found led them to rebrand as an implant center. I also pushed for mapping how the whole company works. I interviewed people on every team, and it turned into a 100+ page guide that new developers used to get started.

I was relocating that year, so I freelanced and built web projects for small clients. One of them, a dental practice’s website, is in Projects below.

Bontech builds data and ML systems for industrial clients. My biggest project was for a gas utility that was supplying gas it never got paid for. I built unsupervised anomaly detection that screened all 890K of their customers. When it flagged someone and the utility sent an inspector, 1 in 8 of those visits found a violation, like a tampered meter or an illegal bypass.

Getting there took a lot of cleanup. The billing data came as an Oracle export: 40 tables, 49.5M bills, and some key fields 69% empty. I cleaned it and moved it into PostgreSQL. Then I built time-series clustering that took the demand model’s R² from 0.60 to 0.85 on test data. The utility’s engineers and I wrote a journal paper about it.

In my second year I was promoted to lead the company’s MLOps project. With a team of three, we reviewed more than 10 vendors first. Then we built our own platform on Docker, K3s, Airflow and Kubeflow, so we could train and deploy models inside client networks with no internet access.

A paid Telegram bot for small businesses. It messages you the moment a payment lands in any of your crypto exchange accounts, so you don’t have to keep checking them.

A Claude skill I made because coding agents talk too much. It makes their answers shorter and easier to read, which helps most when you’re new to a codebase.

I wanted to switch storage backends (S3, MinIO, Backblaze, whatever) without rewriting code or learning a new SDK every time. So I’m building one interface that covers most of the open-source providers.

I didn’t really understand how Python’s event loop worked, so I built something that made me learn. It catches blocking calls inside FastAPI and Starlette handlers and tells you which request caused them. That kind of bug stays quiet until your server falls over under load.

AI coding agents write it a lot, too. When GPT-4.1 got loopguard’s test feedback, its pass rate went from 40% to 98%.

I built this for Rasa, the production system at Mach Milling. The backend uses CQRS, so I needed an inbox/outbox on top of Redis Streams that I could trust. Once it was working in Rasa, I pulled it out into its own library.

Same idea as BucketBridge, but for identity providers. You write against one interface and swap the provider underneath without touching the rest of your code.

A website for a cosmetic dentistry practice. I did all of it myself: the visual design, the frontend and the backend. The point was to help patients who are considering cosmetic work feel they can trust the practice.

Part two, picking up statistical inference where part one left off.

Part one of a teaching video I recorded on the basics of statistical inference.

I wrote this with the gas utility’s engineers, based on the demand modelling we did together at Bontech.