AI systems · backend · reliability · occasionally unnecessary infrastructure
Core Engineer @ Analytics Depot · Building Replex
GitHub · LinkedIn · X · Discord · Email
I build around models that are extremely confident about being wrong.
Most of my time goes into AI agents, backend systems, evals, retrieval, data pipelines, async workers, and figuring out why something that worked perfectly five minutes ago has suddenly developed free will.
Currently Tokenmaxxing @Analytics Depot. Outside work I build whatever rabbit hole looks interesting enough.
| Languages |
 ·
 ·
 ·
SQL
|
| Backend & Systems |
 ·
 ·
 ·
 ·
|
| AI Systems |
 ·
 ·
 ·
 ·
|
| Cloud & Infrastructure |
 ·
 ·
 ·
 ·
 ·
|
| Frontend & Data |
 ·
 ·
 ·
|
Replex status: active
what if making product demos didn't involve fighting a video editor for 3 hours
browser capture → canonical project state → bounded agent edits → verified render
receipts: scenes backed by runnable checks, recapture scoped to what changed, evidence attached to every render
what's actually in there
- Playwright flow + feature brief in, reproducible release video out
- recaptures only the affected scene when the product changes, not the whole video
- currently a local POC with disposable evidence for every render — the unglamorous part that makes agent output trustable
GrabPic status: stable
upload selfie. find yourself in 5,000 event photos. preferably before the event ends.
event isolation, async processing, vector search, idempotent callbacks, deletion flows and enough infrastructure for what originally sounded like "just match faces lol"
receipts: selfie to personal gallery in under 5 seconds, 512-d embeddings, event-isolated search
what's actually in there
- one selfie in, personalized gallery back in under 5 seconds
- 512-dimensional face embeddings with event-isolated vector search
- signed uploads, protected delivery, automated cleanup — the boring parts that decide whether it survives a real event
- if an LLM says it succeeded, verify it
- if something can retry, it will retry at the worst possible time
- "works locally" is not a deployment strategy
- caches are great until they confidently serve yesterday's truth
- async code is extremely fun until ownership becomes philosophical
- users will absolutely find the state transition you forgot existed
- every "temporary" workaround is applying for permanent residency
- taught an LLM to analyze datasets (DataLens)
- built a tiny SOC agent (MicroSOC)
- made deepfakes somebody else's problem (Agentic-DeepFake-Classifier)
- built computer-vision traffic systems (ANPR-and-ATCC-for-Smart-Traffic-Management)
- accidentally built several RAG systems (Permission_Aware_Internal_Knowledge_Assistant)
- taught reinforcement learning to care about derivatives because apparently finance wasn't complicated enough (Derivative_Hedging_Using_RL)
- day job: Core Engineer @ Analytics Depot, building Replex
- after hours: side quests above, quality varies with curiosity
- best way to reach me: email, linked up top
- currently not: opening Kdenlive