Autonomous agents · Research
Multi-Agent Trading Simulator
Four independent AI agents (equities, options, futures, and forex), each running its own budget, strategy, and risk rules. They scan markets on a two-minute loop, open and close their own positions, and write every decision to a live ledger. Each agent keeps a per-instrument track record that feeds back into what it's willing to trade next, so a losing pattern gets down-weighted without anyone touching it.
4autonomous agents
24/7unattended uptime
2mindecision loop
Simulated capital, by design. This is a research rig for agent autonomy: how agents behave over months without a human, how they recover from their own bad decisions, and how they self-correct. It is not a financial product, and we don't sell trading advice.
What it proves: long-running agents that survive real-world failure (API outages, bad data, their own mistakes) and keep operating correctly for months unattended.
Data engineering · Lead intelligence
Public-Records Contact Resolution
A pipeline that turns a business entity into a real human you can actually reach. Commercial skip-trace services return nothing useful when the owner of record is an LLC, so we built the chain ourselves: identify the entity across state registries, extract the officer behind it, then resolve that person to current contact details. Multi-state, fully automated, and built on lawful public record.
2,746leads in one batch
562verified phone numbers
64%entity resolution rate
Compliance built in, not bolted on. Do-not-call scrubbing, opt-out handling, and TCPA-aware outreach rules are part of the pipeline, because the data gathering and the contacting are two very different legal questions.
What it proves: data engineering where the easy path fails: sources that fight automation, records that need real parsing, and results measured in verified contacts, not rows scraped.
RAG · Knowledge graph
Operational Memory Graph
Every engineering session we run gets ingested into a searchable knowledge base: full-text and semantic search over a quarter-million messages, AI-written summaries, and a visual graph of how the work connects. It maintains itself: a nightly job ingests, summarizes, embeds, and tiers older data to cold storage without anyone touching it.
212K+messages indexed
65,924vector embeddings
~$1.40to embed the corpus
Runs on edge inference, not per-token API bills. Summarization and embeddings moved onto Cloudflare Workers AI, which is why re-embedding the entire corpus costs about the price of a coffee instead of a monthly invoice.
What it proves: the retrieval stack behind "search your company's knowledge": embeddings, hybrid search, summarization, and storage tiering, built and running at real scale, not in a notebook.
Infrastructure · Data acquisition
Resilient Acquisition Engine
The data layer under our lead-gen and pricing products. Jobs are declared in config, not code; each one escalates through progressively heavier fetch strategies only when a source pushes back, governs its own request rate per domain, routes through rotating egress, and keeps dedupe state so a monitor run only surfaces what's genuinely new.
30markets, daily
1,406live records maintained
87tests in the suite
Cheap on purpose. The overwhelming majority of runs complete on the lightest tier against public and government sources, which keeps a full nationwide sweep under a few dollars a month in egress.
What it proves: collection infrastructure that survives contact with the real internet (rate limits, shape changes, and flaky sources) and stays maintainable because the jobs are declarative.
Conversational AI · Airline operations
WhatsApp Customer-Service Agent
Built to answer the question airlines lose customers over: what happens when nobody replies? It handles cancellations, refunds, baggage claims and disruptions over WhatsApp, and every incoming message is classified in real time: the right service desk, how urgent it is, how the customer feels, and whether there's legal or regulatory risk in the wording. Anything it shouldn't answer alone opens a tracked case and pulls in a human, who can take the conversation over mid-thread.
8service desks routed
50+intents classified
3languages handled
Born from a real failure. We spent two days trying to cancel a flight, got ignored, and heard back three days later, after the refund window closed. This is that experience rebuilt as software that answers in seconds.
What it proves: an AI agent that knows the limits of its own authority: it resolves what it can, and escalates what it shouldn't touch, with the full conversation handed to the person taking over.
Open source · Creator analytics
Social Analytics Command Center
Built to answer a question every creator gets asked by brands: prove your audience. It scrapes a creator's entire back catalog across Instagram, TikTok, YouTube and Threads, then reports what actually performs, when to post, and which comments deserve a reply. A brand-deals tab tracks the partnership pipeline, and a content analyzer takes any video and pulls frames plus a transcript so you can see why a post landed.
4platforms unified
2.0Kposts analyzed
11.2Mviews accounted for
Run it on your own numbers
- Clone the repo and run
pnpm install
- Drop your exported platform data into
data/
pnpm build, then host the static output anywhere
View the repository on GitHub ↗
What it proves: a full data product end to end, collection through analysis to a shipped interface, and one we were willing to open source rather than describe.
Client pilot · Concluded
Bilingual WhatsApp Booking Agent
A WhatsApp receptionist built for a wellness studio abroad, where WhatsApp, not phone or email, is how customers actually book. It detects which of two languages a customer is writing in and replies in kind, answers from the studio's full service and pricing catalog, hands off into their existing booking system, and escalates anything it shouldn't handle to a human with the conversation attached.
2languages, auto-detected
30+services in catalog
1tap to human takeover
Ran in production with real customers. A pilot engagement that has since concluded. Included here because it's the pattern most service businesses ask us for: the channel your customers already use, answered instantly, with a human always one tap away.
What it proves: conversational AI on the channel that actually converts, with language handling and human handoff designed in from the start rather than bolted on.
