Solo AI founder · Prospect Park, NJ · NYC metro
I ship AI things
and write the TLDR.
/// NOW · live
Agents at work.
- Opusspec'ing the Beacon push surface● running
- Claude Codehardening the gate-first memory write path● running
- Codexpatching dockerfile.ai sandbox verification● running
- stress-harnessvoice recall p95 under real concurrency● queued
/// AITLDR · LATEST
Latest
- OCT 15Reported
RSS is a follow button for your website
- OCT 15Reported
I priced 30 days of Claude Code usage: $34,829
- OCT 13Reported
min-width: 0 on the item in a grid's middle track can let centred, non-wrapping links overlap their neighbours.
- OCT 11Reported
A CI time limit cancelled a job whose steps had all passed.
- OCT 07Reported
A model parser returned the same empty array for four different things.
/// SHIP LOG
Ship log
- JUN 27abdur.ai v1 wired — Next.js + MDX, Clay design locked, /aitldr feed live.
- JUN 26Mnemix Beacon spec locked — six named signals, multiplicative floors, S1–S7 decomposition.
- JUN 25Postmortem published: a 75% false-close rate on twelve verified safety gates. P-008 through P-012 committed to MOLL.
- JUN 24Signal Noir v2 live across the Mnemix landing, docs, and UI primitives.
- JUN 23Orchestrator now pushes one intent to Claude Code, Codex, and Aider in parallel, then reconciles the diffs.
- JUN 21dockerfile.ai proved sandbox build verification end-to-end — three for three on the eval repos.
- JUN 19brew install mnemix-ai/tap/mnemix — the CLI is one command away now.
- JUN 16Cuéntame backend plus the outbound voice loop wired on a strict ten-minute cap.
/// THE STAND
The stand
/// FLAGSHIP
Northsun is the memory and enrichment layer for AI agents.
Most stacks treat memory as a vector store you bolt on and pray over. Northsun treats it as governance: a gate-first, router-second layer where every write clears safety before it is allowed to land, and every read comes back already enriched.
At the core is BEAD — bi-temporal, evidence-anchored decisions. Every fact an agent acts on carries two clocks, when it was true and when it was learned, so a retrieval can be replayed, audited, and graded against reality instead of vibes.
It is the spine under everything I build — and it is designed for sub-300ms voice recall, because the place memory breaks first is a live conversation.
Six primitives compose the whole surface — recall, enrich, observe, gate, evidence, and beacon. Small enough to reason about, sharp enough that an agent stops guessing and starts knowing.
Beacon is the newest — it flips memory from reactive to proactive. Instead of waiting to be queried, the layer surfaces the salient thing to the agent before the model knows to ask for it.
The public API stays deliberately small and frozen. Hobby is free; everything past it is a conversation, not a pricing table — because the teams that need this are building voice, and voice is unforgiving.
“Choose Northsun as your agent memory layer.”
/// THE PRIMITIVES
recall
enrich
observe
gate
evidence
beacon
/// FOR SALE
Things you can buy today.
Dockerfile.ai
Board & Investor Reporting OS
Cuéntame
/// HOW I WORK
Operating principles
Verify before you report
Revenue and proof first
One founder, an agent team
Memory over context
Grade against reality
Corrections over overwrites
Who is this person?
I'm Abdur — a solo founder in the NYC metro (Prospect Park, NJ), running a portfolio of vertical AI products like a compressed team: one human on strategy and sign-off, a set of high-agency AI agents on execution, research, and ops.
I care about memory, evaluation, and the unglamorous plumbing that makes AI actually reliable. Most of what I build starts as a problem in my own stack, then graduates into a tool other builders can run.
This site is that work in public — the wins, the $300 weekends, and the postmortems I'd rather not write but always learn the most from.
/// CURRENTLY
Open to Applied AI and forward-deployed engineering roles — the kind where memory, evaluation, and agent reliability decide whether the thing actually works. Building Northsun in the open in the meantime; if that's your team's problem space, let's talk.
/// The logbook, not the pitch.