Insta-Apply is a working system, not a concept deck. It watches the job market continuously, builds an evidence-cited dossier on any company in minutes, assembles the full application packet, and submits applications from the member's own browser under a set of laws enforced in code. A cast of voice-native AI headhunters onboards the member by conversation, files every confirmed answer into a private profile graph, prepares them for interviews, and refuses to invent a fact it does not have. The platform itself was built the same way it runs: a fleet of coding agents working in parallel lanes, each piece landing only after deterministic validation. Every number in this paper was measured on the running system in August 2026.
§1System map
Six layers, strictly ordered. Data flows down the stack; nothing reaches a submitted application without clearing the gate layer; the ledger under all of it records everything once and never rewrites it.
L1Discovery ledger
Continuous sweeps of the monitored job boards, one immutable record per observation: 28,080 sweeps between 4 and 31 August 2026. From this the system computes hiring velocity, role-open medians, and closure rates per company, the raw material for every recommendation. A role that closed while the member watched is a fact the member deserves to know before spending an evening on it.
L2The member graph
One private graph per member holds every fact the system may use: profile answers, work history, story bank, preferences, languages. Writes are confirm-gated by law: the AI proposes, the member confirms, only then the graph changes. A regulated answer (veteran status, disability, demographics) that the member never stated is never invented; the question rides pending and the gate holds the application instead of guessing.
L3Packet factory
Per job, the factory assembles the tailored resume, the cover letter, the answer set, and the dossier summary. A packet is complete or it is named partial with the reason recorded; a partial packet cannot be submitted, and the submit path reports exactly which artifact is missing. Mean measured build: 148.5 seconds over 170 historical builds.
L4Apply automation, in the member's browser
Submissions run in a browser extension on the member's own machine and their own sessions, not from a server farm that job boards can fingerprint. Eight job-board adapters are covered by a single gate suite (64 rows, 7 mutants, zero survivors). The gate laws are enforced in code: no submit without a complete packet, no invented regulated answer, and every terminal path of a fill carries an outcome and a sentence, so a fill that never submitted can never look like one that did.
L5Voice-native AI headhunters
A cast of personas (American, Chinese, Indian, and Latin voices; the member's practice language drives the accent matrix) leads onboarding by conversation and files confirmed answers into the graph. The voice loop is a four-state, eight-event machine with illegal transitions refused by name; barge-in runs on a real voice-activity detector, not microphone energy. Turns are budgeted under 1.5 seconds to first token; the measured p50 on the routed model is 725ms.
L6Event ledger and metrics
Everything the system does is one immutable event in an append-only, per-member ledger with deterministic ids: a retried write produces the same id and counts once, so conversion numbers stay honest under retries. The taxonomy is closed; an event carrying member content (an answer, a transcript, a message body) is refused by the ledger itself. Every metric on the operator dashboard is a deterministic rollup over this file, recountable forever.
§2Speed and capability, measured
Every figure below was measured on the running system, not quoted from a vendor page. Where a figure is an estimate the table says so and names its basis.
Latency
Source: voice-lane latency probes, Oracle host, 30–31 Aug 2026. The 9.7-second full turn decomposed into speech recognition, reasoning, and speech synthesis; the reasoning leg was the dominant term, which is what the model reroute removed.
Throughput and scale
| Capability | Measured value | Basis |
|---|---|---|
| Market observation | 28,080 sweeps in 27 days; roughly 1,040 per day sustained | Discovery ledger, 4–31 Aug 2026 |
| Company dossier build | 148.5s mean; 582 per day server-side | 170 historical builds; interval sweep over 243 builds found zero overlap |
| Application fill | Eight job-board adapters, one gate suite, 64 rows green | Submit-outcome gate, 31 Aug 2026 |
| Connection ceiling | 0.049% errors at 2,000 concurrent; 0.086% at 5,000 | Load battery after the descriptor-limit fix (the wall was 724 before it) |
| Crash recovery | Automatic restart, both services | systemd, verified live 31 Aug 2026 |
§3The build itself: a managed fleet of twelve agents
Insta-Apply was built by a fleet of twelve coding agents running continuously, coordinated by a managing agent under two disciplines this project names loop engineering and graph engineering.
Loop engineering. Every agent works from a written order in a shared ledger; no work merges to the main branch until the managing agent runs the suites itself; every deploy is verified against the live site, not the build log. The loop runs day and night, and its state lives in an append-only record so a restart or a context reset loses nothing. Agents that finish get the next hour of work from the ledger; agents that stall get found by the same record, not by memory.
Graph engineering. The product's core data structure is a closed graph: member facts, questions, answers, jobs, companies, and events are nodes with named edges, and the vocabularies are closed so two parts of the system cannot drift into private dialects. The same discipline governs the build: the shared order ledger is itself a small graph of claims, each stamped, each auditable, each survivable.
Geometric reasoning, in practice. The fleet treats the codebase as a shape with invariants rather than a pile of files: a seam must have a producer and a consumer that meet; a taxonomy must have exactly two lists that stay identical; a route count must cover exactly the public surface. Six production defects in one night traced to one geometric failure, two halves built and nothing checking they met. Every one of those invariants is now an executable probe.
Recursive self-improvement, with evidence. The fleet's failure modes become its own test battery. An alarm that compared two copies of the same wrong answer was rebuilt to call the code instead of read the text; a harness that printed failure about a product it had never driven was made to drive it; a fabricated session fixture was deleted and its absence pinned by a row that kills a mutant reintroducing it. In one week the suites caught fourteen defects in the agents' own work, including four in a single lane's output. That conversion loop, incident to permanent guard, is the moat that compounds.
§4The AI agent: capability and cost, honestly stated
The member-facing agent does four jobs: onboarding by conversation, answering application questions, interview preparation with scoring, and negotiation coaching. It speaks in a cast of personas whose voices and accents follow the member's practice language, and it refuses off-topic work in character, because the product is a headhunter, not a general chatbot.
| Cost component | Value | Basis |
|---|---|---|
| Application packet (resume, cover, answers, dossier) | $0.087 metered | Ledger cost events, post-reroute; 97% of prior cost traced to one vendor |
| Reasoning per voice turn | sub-cent | Routed frontier open-weights model, metered per token |
| Speech recognition per minute | fraction of a cent | Vendor streaming list price, 31 Aug 2026; verify before quoting externally |
| Speech synthesis per turn | about one cent | Vendor per-character list price at typical reply length; estimate, verify before quoting externally |
| Full ten-minute voice session | estimated $0.40 to $0.60 | Composite of the three lines above; estimate with named basis, not a metered figure |
One discipline keeps the agent affordable at scale. Hard caps: every voice session carries a token and dollar budget, and a silence timer disconnects an idle call in the persona's own words rather than billing an open line.
§5The data moat, visualized
What the discovery ledger sees that a job board never shows a candidate: velocity, churn, and which postings are real. One live example from the ledger, a streaming-device company tracked through August 2026:
Source: Insta-Apply discovery ledger, 264 company rows, 28,080 sweeps, 4–31 Aug 2026. A company opening 264 roles and closing 16 is accumulating listings, not hiring at the same rate. The member sees this before writing a word.
Open roles by family, same company, same window
The shape of the opening tells the member where the company's money is actually going. Every dossier carries this table, computed live, with its gaps labeled rather than hidden.
§6What makes it distinctive
A discovery ledger that records the job market itself at sweep grain, 28,080 observations and growing, which no vendor sells; a legal-and-truth gate stack enforced in code where every refusal has a name; a voice-native headhunter layer whose onboarding conversation is also the data-entry layer; a build process run by a twelve-agent fleet whose own failures become its test battery; and a validation culture where every law is a suite, every suite has mutants, and every claim in this paper maps to a commit and a green row.
Road ahead. Doors open 14 September 2026 at the UF AI Day career fair, with recruiter access through NFC and QR credentials. The fair funnel (scan, one-tap signup, avatar onboarding, one free dossier and one automated application) is the first live load test; the metrics surface for partners hangs off the operator's site rather than the product, so the numbers a recruiter sees are the same ledger numbers the engineering team reads.
Insta-Apply is built and operated by Daniel Roberts. Platform figures measured on the running system, August 2026; figures named as estimates carry their basis and a verify-before-quoting note. Companion research: Chimera (autonomous trading research) and Prescient (prediction-market truth estimation).
Contact and live metrics: daniel-roberts.com · Product (opens 14 Sept 2026): insta-apply.com