GIL NETO
Founder-Engineer & Senior Software Engineer, AI Products
Senior engineer, 12 years. I build LLM-backed systems where the model does judgment and deterministic code does correctness. Most of that work now sits in memory: how an agent captures what it learns, where that knowledge lives, how it comes back, and what stops it confidently repeating something that stopped being true. I run my own second brain on those rules, and I ship B2B utilities end-to-end — infrastructure to UI — solo.
Philosophy
→Memory as Infrastructure
An agent that forgets is a demo. I build the layer underneath: capture that cannot be skipped, routing that runs on a schedule rather than on attention, retrieval split between generated indexes for state and vector search for facts, and a nightly eval that says out loud when the index has started lying. Judgment is the model's job. The guarantees are the harness's.
→Infrastructure-First
I own the boundary — from DNS record to React component. Several products in production behind one operational surface: Docker, Nginx, Postgres, durable workflow orchestration, self-hosted analytics, all containerized and reproducible. Runbooks and an infrastructure map kept in sync with prod.
→AI-Augmented Execution
AI coding agents kill boilerplate. That frees me to spend cycles on system architecture, data modeling, and business logic — not scaffolding. The result: solo-shipping multiple products without cutting corners on tests, types, or observability.
Experience
Independent Founder-Engineer
Building and operating B2B utilities solo — infrastructure, backend, UI, billing, ops — plus contract work as a senior full-stack engineer on AI product teams, where the current brief is the knowledge and memory layer an LLM product runs on.
- →Agent memory system in daily use: an LLM agent reading and writing a persistent markdown knowledge base, where capture, routing and checkpointing are enforced by lifecycle hooks instead of prompts, and a scored nightly eval gates the whole thing.
- →Retrieval designed in two rungs — generated indexes for state, vector search for a specific fact — with "not in the corpus" kept as a valid answer rather than a failure to paper over.
- →Kelaro (live, kelaro.io) — accounting-automation SaaS in closed beta with active pilots (PT chartered accountant, NL), powered by Koa.
- →Koa — deterministic document-extraction engine, built as the deliberate alternative to vision-LLM extraction: balances must reconcile, and it refuses loudly rather than returning a plausible total.
- →Multi-product operational topology: Nginx, Docker, Postgres 16, Temporal, transactional and inbound mail. Authored infrastructure map and runbooks.
- →Agentic development pipeline with a decision gate ahead of specification, so agents execute settled decisions rather than improvising undecided ones.
Lead Software Engineer
Promoted to Lead. Drove monolith-to-microservices migration and feature delivery across the enrichment platform while managing a development team.
- →Owned feature delivery for high-traffic enrichment pipelines processing data at production scale.
- →Designed and shipped the LLM-integration layer (prompt orchestration, structured-output extraction) powering production enrichment workflows.
- →Stood up the observability stack (Grafana, CloudWatch, structured logging), reducing incident MTTR.
- →Mentored junior engineers across teams of 4–5. Introduced code-review and CI quality gates that lowered regression rate.
Senior Software Engineer
Joined as Senior to tackle full-stack challenges in a high-paced environment.
- →Built and maintained a scalable design system consumed across multiple product surfaces.
- →Owned front-end architecture decisions (state, routing, build) across feature teams.
- →Drove client-facing projects across sectors (health, retail, others).
Front-end Lead Developer
Built and led the front-end for Glartek's core industrial-IoT product across multiple iterations. Defined the component library, build pipeline, and front-end architecture that supported enterprise deals with industry-leading manufacturing and oil & gas companies.
Analyst / Consultant
Delivered features across multiple enterprise applications. Final two years: on-site systems administrator for a major oil & gas project in Kuwait, owning operations and incident response.
Selected Work
Second Brain
Running DailyPersistent Memory for an LLM Agent
A markdown knowledge base an LLM agent reads and writes as its long-term memory, in use every day since Jan 2026. The prompts were never the hard part. Capture that cannot be skipped, routing that happens on a schedule rather than when someone remembers, retrieval in two rungs (generated indexes for state, vector search for a specific fact), and a scored nightly eval that catches the index lying before I act on it. Two published write-ups, both with the real numbers rather than the flattering ones.
Lifecycle hooks enforce capture and checkpointing outside the model, so a skipped save is structurally impossible rather than discouraged. A nightly batch routes the inbox, regenerates every index, runs a scored eval against a fixed question set, and audits the previous day's work for drift. Context reads are capped and the cost per night is a tracked number.
Kelaro
Live · Closed BetaAccounting-Automation SaaS
B2B SaaS for fractional CFOs and accountants. Ingests bank PDFs and Open Banking (PSD2) feeds, runs deterministic extraction via an internal engine (Koa) orchestrated by Temporal, and emits accountant-ready datasets. The model may explain a number. It never produces one. Active pilots: PT chartered accountant + NL.
Next.js + Postgres 16 + Temporal workflows. Stripe integration with founder-discount pipeline (coupons, promo codes, 30-day trial). Admin dashboard for Users / Waitlist / Promo Codes / MRR. OAuth 2.0 against GCP and Azure Entra ID enterprise tenants.
Koa
EngineDeterministic Document Extraction — the Alternative to Vision-LLM
Config-driven extraction system that turns bank-statement PDFs into structured JSON. Built deliberately against the vision-LLM approach: balances have to reconcile and the engine refuses loudly when they don't, rather than returning a plausible total. A confidence score is not a proof. Three components: a Python/FastAPI engine, a Next.js + Prisma mapper UI for authoring per-bank templates, and an isolated PII scrub container. Powers Kelaro's extraction pipeline.
Python 3.12 / FastAPI with pdfplumber + pikepdf for native PDFs and Tesseract / PaddleOCR fallback for scans. Pydantic schemas end-to-end. mypy strict, ruff, CI-gated, auto-deployed via GitHub Actions on push to release branch.
Tech Stack
Languages
Frameworks & Runtime
AI & LLM
Infrastructure & Cloud
Integrations
-
The index was lying and the eval knew
Six weeks after I fixed capture with hooks, my second brain was still confidently wrong. The nightly test had been saying so for a month. Nobody read the number.
-
My agent had two rules and obeyed neither
Both rules were well written. They contradicted each other and I hadn't noticed, because writing a rule and making a decision are not the same job.
-
Running a second brain on an LLM agent that forgets
I told the agent to save my context. Sometimes it did. The fix wasn't a better prompt - it was hooks, structure, and economics.