Aaron
Altergott

Operations engineer building with cloud and AI. If a workflow is slow or doesn't have a good solution yet, I build one.

About

AWS Certified operations specialist who builds tools to make life easier. Worked across customer ops, marketing, finance, sales, and engineering which gives me a different perspective than someone who came up through one track.

I've automated workflows, fixed recurring failures, and tightened up processes that were wasting time and money at work. That is what pulled me towards tech.

I enjoy figuring out why something isn't working and making it work, or taking something that kind of works and making it more efficient. Cloud and AI is where that clicks for me. They overlap enough that you can't really work on one without running into the other.

Skills

Cloud

AWS Bedrock · Lambda · API Gateway · S3 · CloudFront · IAM · CloudWatch · Serverless · AWS CLI

AI / ML

RAG · Vector DB (Pinecone, FAISS) · Embeddings · HDBSCAN Clustering · LSH MinHash · Semantic Search

Backend

Python · FastAPI · REST APIs · SQL · SQLite · CI/CD

Tools

Claude Code · GitHub · Docker · Codex · Linux · PowerShell · Notion · Vercel · Supabase · Sitelink ERP

Ops

Customer & tech support · Financial reporting · Competitive analysis · Client outreach · Accounts Payable

Experience

Nov 2024 – PresentSouthwest Self StorageRedlands, CA

Operations Specialist (CSR)

Eliminated 60+ monthly AMEX and Discover autopay failures by tracing them to an AVS misconfiguration
Cut auction mail certification processing time by 90% with an automated workflow using file conversion and AI-assisted parsing; authored the company process document
Reduced average monthly past-due balances by 13% after configuring automatic SMS to overdue accounts
Reconcile daily financial reports in ERP; monitor AP for duplicates, fraud, and errors before payment
Run customer support for 500 tenants, handling inquiries and account issues through to resolution
Analyze competitors and adjust weekly pricing and promotions to stay competitive while maintaining monthly KPIs

May 2026 – PresentNull404Remote

Operations Volunteer

Lead sponsor outreach for a cybersecurity research group, secured 2 gear sponsorships
Designed the Null404 sponsorship package PDF; assist in development of CTF challenges

May 2023 – Aug 2023The Advance GroupRemote

Intern

Analyzed voter records across NYC districts to identify demographic patterns and developed targeted policy recommendations for a councilmember
Vetted ~100 NYC organizations across industries and wrote partnership proposals that shaped client outreach decisions

Projects

Orca active

Orchestration runtime that routes each part of a goal to the right agent, holds each to a contract, and returns a merge recommendation with the evidence attached.

Delegates each task to the right agent automatically: Claude handles implementation, Codex handles review, and your CI gates run verification. No copy-pasting between windows, no re-explaining context
Before any agent runs, Orca translates the goal into specific, checkable obligations. Conditions like "tests in internal/store must pass" are defined upfront so every agent works against a contract, not an open-ended instruction
Each agent run is isolated in an Execution Capsule with exactly the files it needs, a token budget, and verification gates; a patch only advances when the gates clear and evidence is attached

Source ↗

goal plan capsule verify merge

The problem

An agent that says tests pass is not the same as one that shows you the logs. Most agent-assisted workflows collapse at that gap; there is no way to know if the agent actually ran the tests or just said it did. Orca defines what done means upfront, runs the gates, and stores the artifacts that prove it.

Decisions

Exit conditions are defined before any agent runs, not negotiated after. An agent working against a contract cannot quietly decide it is done when it is not.
Each agent gets a briefing compiled from current goal state, not a replay of the full conversation. Claude does not need Codex's review notes, and Codex does not need Claude's implementation history; keeping them separate is what keeps each agent on task.
Every step is written to the event log before it runs, so a killed process always has a checkpoint. Reconstructing state from memory after a crash was too unreliable to trust.

Claude Chaperone complete

Installs a plan-build-audit-commit workflow into any Claude Code project with a single command.

Twelve slash commands step through planning, phase splitting, test-first building, diff auditing, and commit — each stage runs in its own cleared context so Claude does not carry assumptions from prior work
Four stdlib Python hooks handle scope drift warnings at end-of-turn, git push confirmation prompts, build log sync reminders, and session state injection after /clear so Claude picks up where it left off
The installer merges settings.json without clobbering existing hooks and runs a 58-test suite to confirm the install is clean; idempotent and safe to re-run on existing projects

Source ↗

plan build audit commit

The problem

Claude accumulates issues in long sessions by carrying assumptions, filling in gaps it shouldn't, and drifting from scope. Mandatory /clear between every stage fixes this as each phase runs in a fresh context. session_start.py injects previous state after /clear so the thread doesn't break.

Decisions

A one-file fix or a typo doesn't need this workflow, only scoped for multi-phase work. The overhead only pays off when the task spans multiple sessions.
If codex, gemini, or aider are on PATH, they run the audit pass. If none are present, an isolated fresh subagent handles it.
Three loop cap on /re-audit keeps the workflow from turning into its own kind of scope drift. /handoff writes a self-contained state file mid workflow and the next session reads it and picks up without you re-explaining what happened.

Narrative Intelligence Engine complete

Tracks financial narratives across news, SEC filings, and Reddit, scores their momentum, and maps them to S&P 500 tickers.

Deduplicates incoming documents with LSH MinHash (0.85 Jaccard threshold), clusters survivors into narratives using HDBSCAN, and tracks centroid drift over time as a signal of narrative mutation
Maps narratives to S&P 500 tickers by computing cosine similarity between narrative embeddings and a pre-built library of ticker embeddings generated from SEC 10-K summaries
Flags coordinated campaigns: five or more sources posting about the same narrative within 300 seconds triggers the adversarial burst detector

Source ↗

sources cluster score map

The problem

Financial narratives are scattered, so the same story surfaces as an RSS headline, a Reddit thread, and an SEC filing. NIE aggregates those fragments across sources every four hours, clusters them into tracked narratives, and watches how they evolve.

Decisions

The number of active narratives at any moment isn't knowable in advance, HDBSCAN discovers it in the data.
When a narrative cluster's center shifts semantically, that movement is what gets flagged, not a keyword change or title match.
Emerging, Growing, Mature, Declining, Dormant lifecycle stages. A dormant narrative that spikes past 0.10 velocity reverts to Growing automatically because going quiet and then spiking is itself a signal.
The 300-second window in the coordination detector is short enough to catch a coordinated burst, long enough to filter organic spread across sources. False positive filtering is straight forward when sent to Sonnet for analysis.

AWS Docs RAG complete

Answers questions about AWS services by retrieving relevant documentation and generating grounded responses with Claude.

Scraped about 120 pages across five AWS services, split into 1000-character chunks with 200-character overlap, embedded with Amazon Titan v2 (1024 dimensions), and indexed in Pinecone with cosine similarity
At query time: embeds the question with the same Titan model, retrieves the top 5 chunks from Pinecone, assembles a grounded prompt with anti-hallucination instructions, calls Claude via Bedrock, and returns the answer with source URLs
Each ingestion script writes a manifest and checks that the previous stage completed before running, so partial failures do not corrupt the index

Source ↗ Live ↗

docs chunks embed retrieve answer

The problem

Claude can describe AWS services confidently from training data, which is precisely the issue, training data isn't the current documentation. This RAG retrieves the relevant chunks from the actual docs at query time and builds the prompt around them. Responses come back grounded in reality with source URLs attached. I used this to study for my AWS CP exam.

Decisions

S3, EC2, Lambda, DynamoDB, and VPC cover what shows up in almost every AWS architecture question, questions about one usually involve another.
A concept that starts near the end of one chunk appears at the start of the next because of the 200 character chunk overlap. Retrieval does not miss it because of an arbitrary split boundary.
Pinecone free tier over OpenSearch Serverless keeps it at ~$3/month vs ~$700. Same architecture but that price difference is what made this easily deployable.
Adding new AWS services is as easy as putting the URLs in the first ingestion script and triggering a rerun.

Contact

Connect with me here:

Emailadaltergott@gmail.com

GitHubmicronwave

LinkedInaaronaltergott

LocationRedlands, CA

Download Resume ↓

AaronAltergott

About

Skills

Experience

Operations Specialist (CSR)

Operations Volunteer

Intern

Projects

Contact

Aaron
Altergott