A finance operations workspace that investigates invoice exceptions across original records and produces a cited review packet before anyone changes the books.

Built for Open Agent Hackathon 2026, Track 1: Solving Fragmented Intelligence, with NVIDIA Nemotron as the core investigation model.

Links for judges

Resource	What it shows
Live app	The deployed UnitShift workspace, where you can run a real investigation
Demo video (3:07)	A narrated walkthrough, including a live investigation at 1× speed
Pitch deck	The problem, agent workflow, verification, sponsor technology and impact
Submission page	The video, pitch deck and captions in one place
Recorded investigation	A recorded live run of the featured cancellation case
Call receipts	Eleven hash-linked records of every source query and model call
Semantic SQL · Registration script	How the Zetaris source tables and join are registered
Data manifest	Data attribution, source hashes and transformations
Verification record · CI	Test and deployment evidence on Windows, macOS and Linux
Deployment notes	How the judge deployment is hosted and its current limits
Media provenance	How the video, voice, music and thumbnail were made, with licenses
License · Third-party notices	MIT for UnitShift code, plus upstream and dataset licenses

The problem

An invoice exception rarely contains its own explanation. A reviewer may have to reconstruct the customer context, inspect other transaction lines, compare product identifiers and check reporting-currency references. A convincing model explanation is not enough when the original invoice relationship or payment authorisation is missing.

UnitShift keeps that investigation in one workspace. It separates what the source records, what the model infers, what code verifies and what evidence the reviewer still needs.

A real example

In the published UCI dataset, customer 16446 has two records for stock code 23843 on 9 December 2011:

Source workbook row	Invoice	Quantity	Unit price, GBP	Calculated line value, GBP	Recorded time
540423	581483	80,995	2.08	168,469.60	09:15
540424	C581484	−80,995	2.08	−168,469.60	09:27

The lines add to zero. Their matching attributes support a candidate relationship. They do not establish the commercial reason, original-invoice reference, payment settlement or refund approval. UnitShift identifies the candidate and requests those missing documents instead of treating arithmetic as permission to act.

What works

Workspace	Working capability
Overview	Source coverage, historical context and direct investigation entry
Review queue	Real cancellation, duplicate-looking, missing-identity and zero-price signals
Invoice ledger	All 541,909 original lines, exact invoice/customer/SKU search, original workbook row numbers
Customer records	Published anonymised identifiers and transaction context
Product signals	Source descriptions and description variants without invented inventory facts
Investigation lab	Live source reads, bounded model tools, typed decisions, verification feedback and a visible evidence gate
Sources & evidence	Publisher credits, reuse terms, source hashes, registered tables and semantic SQL
Business model	An editable pilot-value calculator with explicit assumptions

A reviewer can inspect actual agent steps and query/model receipts, export JSON or a readable review packet, and save an approved evidence check in scoped memory. Progress indicators follow actual receipts. A failed live request is shown as a failure; a saved result is labelled as a recorded run.

The earlier stock-recovery experiment remains separately labelled at /sandbox. Its commercial records are synthetic. It is read-only on the public deployment and documented in sandbox notes.

Architecture

flowchart LR
    subgraph Data["1 · Attributed sources"]
        direction TB
        UCI["UCI Online Retail<br/>541,909 original lines"] --> Z["Zetaris semantic view<br/>4 tables · no fan-out"]
        ECB["ECB GBP/EUR rates<br/>272 observations"] --> Z
    end
    subgraph Agent["2 · Bounded investigation"]
        direction TB
        M["H-MEM<br/>approved checks"] --> N["NVIDIA Nemotron 3 Super<br/>on Nebius"]
        N <--> T["Allow-listed read tools"]
        N --> C["Code checks<br/>rows · time · sum · FX"]
        C --> G["Policy evidence gate"]
        G --> NS["Nightshift<br/>second NVIDIA review"]
    end
    subgraph Out["3 · Human decision"]
        direction TB
        P["Cited review packet<br/>no money moved"] --> R["Reviewer approves<br/>or requests evidence"]
    end
    Data --> Agent --> Out

Loading

Hosting: CloudFront HTTPS in front of an AWS EC2 host that runs the app, source reads, H-MEM and Nemotron calls. Nightshift runs on a Mac through a private loopback SSH bridge. A reviewer-approved check is saved back to H-MEM for later recall.

Technical implementation

• Source grain stays intact. The importer preserves original workbook row numbers, missing identifiers and duplicate-looking records. It derives invoice/product summaries without presenting them as separate publishers.
• Zetaris performs the actual join. unitshift_control.retail_evidence joins the transaction grain, invoice summary, product catalogue and dated ECB observation. The recorded full join count is 541,909, so the view does not multiply the source grain.
• Nemotron uses a bounded tool contract. It selects from permitted read tools and returns a typed, source-cited proposal. The application checks cited IDs, relationship fields, chronology, arithmetic and FX direction. Invalid proposals can receive verification feedback for repair.
• Policy requirements remain independent. Mandatory missing evidence is imposed even if the model omits it. The raw proposal and policy_added_missing_evidence remain distinguishable in the result and receipts.
• Memory requires operator approval. H-MEM stores reusable checks with a dataset/contract scope. A stale scope cannot silently become the current check.
• The audit trail is inspectable. Receipts contain query/model identities, token use, input/output hashes and linked record hashes. This detects changes within the supplied chain; it is not external notarisation.
• Live reads fail closed. An unavailable or changed source does not trigger a fixture answer. Jobs persist in SQLite; interrupted jobs are marked explicitly. Public calls are queued and rate-limited.

Sponsor technology

Technology	Actual role	Proof to inspect
NVIDIA Nemotron 3 Super on Nebius Token Factory	Source-tool selection and cited investigation decisions	Effective model nvidia/nemotron-3-super-120b-a12b, actual call usage and verified proposals
Zetaris	Source verification and the semantic join across retailer context and independent currency reference	Registered source counts, SQL, full-grain readback and query receipts
Meterless Nightshift	A separate structured request-for-evidence review through a working NVIDIA provider	Recorded effective model nvidia/nemotron-3.5-bolt-e4b, route evidence and accepted review
Meterless H-MEM	Approved evidence-check storage and scope-bound recall	Pinned upstream reference engine, ledger IDs, mem-1 recall in the recorded run and portability integration checks

The recorded public run completed in 10.89 seconds and includes all four integrations; the live run captured in the demo video took 11.99 seconds. That is one observed run, not a latency guarantee or broad accuracy result. Model names are recorded effective identities; availability depends on the provider account.

Real data and permitted reuse

Publisher	Data used	Attribution and reuse
UCI Machine Learning Repository	All 541,909 Online Retail transaction lines, 2010–2011	Daqing Chen (2015), DOI 10.24432/C5BW33, CC BY 4.0
European Central Bank	272 daily GBP-per-EUR reference observations	ECB data API, statistics reuse policy

Raw ECB observations are unchanged. Reporting equivalents are UnitShift calculations using GBP divided by GBP per EUR, with the same or preceding source calendar date. They are retrospective reference conversions, not settlement or point-in-time trading rates. The source does not specify a timezone, so source calendar timestamps are preserved.

The review queue is a disclosed selection of 40 value-ranked source signals, not a balanced evaluation sample. Identities remain anonymised as published. The source manifest records hashes, transformations and selection rules. Neither publisher endorses UnitShift.

Run locally on Windows, macOS or Linux

Requirements

• Python 3.11+, installed and available through python3, python or the Windows py launcher.
• Node.js 18+ and npm. CI uses Node.js 22 and Python 3.12.
• Git and an internet connection for the initial dependency and public-data downloads.
• Enough free disk space for the original archive, CSV exports and SQLite index. No local GPU or model download is needed.

Python application code uses the standard library. Node runs the pinned H-MEM TypeScript reference engine. AWS CLI is only needed if you choose to register your own S3 source deployment.

Quick start: same commands on all three operating systems

Run these in Terminal, a Linux shell, or Windows PowerShell:

git clone https://github.com/TusharTechs/unitshift.git
cd unitshift
npm ci
npm run setup:memory
npm run setup:env
npm run setup:public
npm start

Open http://127.0.0.1:8787.

setup:env creates .env.local only when it does not already exist. setup:public downloads the two primary public sources, checks the UCI archive hash and builds the complete local browsing index. The initial import can take a little time. Downloaded data, generated indexes, credentials and memory state stay in ignored .state/ or .env.local paths.

Local browsing needs no paid API account. A new checkout can inspect the ledger and recorded evidence. A fresh live investigation requires the provider configuration below. The app does not substitute a recorded result for an unconfigured live run.

If Python is installed outside your PATH, set its executable explicitly:

# Windows PowerShell: replace the example path with your actual Python executable.
$env:UNITSHIFT_PYTHON = "C:\path\to\python.exe"
npm start

# macOS or Linux: use your actual Python executable path.
UNITSHIFT_PYTHON=/path/to/python3 npm start

If you already downloaded the source files, import them without downloading again:

npm run import:public -- --retail-zip "/path/to/online-retail.zip" --ecb-csv "/path/to/ecb-gbp-eur.csv"

Enable live investigations

Edit your local .env.local, using your own credentials:

Variable	What to set
ZETARIS_BASE_URL	Origin of your working Zetaris AI Data Gateway
ZETARIS_API_KEY	Gateway API bearer token
ZETARIS_ORG_ID	Numeric organisation ID, not the cluster hostname
NEBIUS_BASE_URL	Your Token Factory OpenAI-compatible API base, such as https://api.tokenfactory.nebius.com/v1
NEBIUS_API_KEY	Your server-side Nebius key
NEMOTRON_MODEL	An available NVIDIA Nemotron model from your account catalogue
NIGHTSHIFT_BASE_URL	The OpenAI-compatible runtime endpoint shown in Nightshift AI Hotspot
NIGHTSHIFT_API_KEY	The runtime connection key shown by Nightshift
NIGHTSHIFT_MODEL	An available runtime model or route, such as meterless/deep
DATA_BACKEND	zetaris for live source-backed investigations

API keys stay server-side and are never needed in the browser. The route label deep does not itself guarantee a larger model; inspect the effective model in the receipt.

Register the generated CSVs in your own S3 bucket and Zetaris organisation. The registration script supports an uploader AWS CLI profile and a separate, temporary, read-only source session:

npm run register:public -- --bucket YOUR_BUCKET --region YOUR_BUCKET_REGION --profile YOUR_UPLOAD_PROFILE --credentials-file ".state/source-session.json" --create-container

The source-session JSON must contain temporary AccessKeyId, SecretAccessKey, SessionToken and Expiration values for a role allowed to read only your source prefix. Never commit it. The uploader profile needs permission to put the four generated files under unitshift/v1/public/ in your bucket. Omit --create-container if unitshift_control already exists. Source credentials expire and must be renewed; the judge host's renewal script is deliberately tied to its dedicated deployment role.

Alternatively, omitting --credentials-file uses temporary credentials exported by AWS CLI v2 from the selected profile. Long-lived IAM user credentials are rejected. Use a least-privilege source role; the script never writes credential-bearing SQL into receipts.

The script creates/updates the four source tables, registers the semantic view and verifies the full 541,909-row count. It does not reuse the original author's bucket by default.

Verify your checkout

npm test
npm run check:platform
node --check static/replay.js
node --check static/workspace.js

npm test runs 48 regression checks. The separate platform integration check exercises the actual Node/H-MEM bridge, approved-check recall, stale-scope rejection and read-only SQLite paths containing spaces. With imported data, it also verifies the complete source count and exact invoice navigation. The CI matrix runs on Windows, macOS and Linux, including the complete public-data setup.

Project structure

unitshift/
├── app/
│   ├── server.py                 HTTP API, static routes and asynchronous jobs
│   ├── replay.py                 Real-data browsing, investigations and evidence gates
│   ├── providers.py              Zetaris, Nebius and Nightshift adapters
│   ├── memory.py                 Cross-platform Node/H-MEM bridge
│   ├── receipts.py               Hash-linked query and model receipts
│   ├── config.py                 Server-side configuration
│   └── operations*.py            Separately labelled recovery sandbox
├── static/                       Evidence workspace, sandbox and engineering UI
├── assets/                       Original UnitShift SVG logo and wordmark
├── data/
│   ├── public/                   Attribution manifest, source snapshot and semantic SQL
│   └── evidence/                 Recorded runs, actual receipts and verification evidence
├── scripts/
│   ├── python.mjs                Python launcher for Windows, macOS and Linux
│   ├── download_public_data.py   Primary-source download and import setup
│   ├── import_public_retail.py   Source-preserving XLSX/CSV ingestion
│   ├── register_public_data.py   Own-account S3/Zetaris source registration
│   ├── memory.ts                 Pinned H-MEM reference-engine calls
│   ├── check_portability.py      Real memory and read-only source integration checks
│   └── refresh_source_access.py  Dedicated judge-host source renewal
├── deploy/                       systemd app and source-renewal units
├── docs/                         Deployment notes, screenshots and sandbox documentation
├── tests/                        Reconciliation, recovery and real-data regression tests
├── .github/workflows/ci.yml       Three-OS verification matrix
├── .env.example                  Configuration names without secrets
├── THIRD_PARTY_NOTICES.md         Upstream and dataset attribution
└── LICENSE                       MIT for original UnitShift code

vendor/hmem/, node_modules/, .env.local and .state/ are generated or private and are excluded from Git. H-MEM is fetched from upstream commit 1742dec56a66e58eef9dda2d2b865823758ca8cd.

Evaluation, business model and current boundaries

The recorded real-data examples cover cancellation context, repeated cancellation rows and missing customer identity. The regression suite checks invalid references, source changes, join fan-out, chronology, FX direction, unsupported approval and public call limits. These demonstrate implemented behavior, not independently measured production accuracy.

The proposed first customers are finance operations teams at wholesalers and multichannel retailers. The £499/workspace/month offer and review-time calculator are pilot hypotheses. A customer pilot would measure review time, incorrect links, rework and willingness to pay. No paying customers, recovered funds or measured savings are claimed.

The public app runs on AWS EC2 with encrypted persistent storage behind CloudFront HTTPS. The primary app, source reads, H-MEM and Nemotron inference run in the cloud. Nightshift currently runs on the author's Mac through a private loopback SSH bridge; if that runtime is unavailable, the secondary review is explicitly unavailable while the primary app remains accessible.

The initial judge compute stop is configured for 10 November 2026, 00:00 UTC. Hosting cost estimates, source renewal and retirement details are in the deployment notes. Production work includes authorised business connectors, tenant authentication, independent evaluation, private/TLS origin transport and a supported unattended Nightshift host. The prototype creates no external orders, messages, refunds or payments.

License and provenance

Original UnitShift code and logo are available under the MIT license, copyright 2026 TusharTechs. That license does not replace the licenses for upstream engines, vendor services, datasets or model weights. H-MEM retains Apache-2.0, UCI data retains CC BY 4.0, and ECB data retains its published reuse terms. See third-party notices and media provenance.

AI disclosure

Area	Tools and contribution
Research and product development	OpenAI Codex assisted research, problem selection, architecture, implementation, debugging, test creation, deployment setup and documentation. Source facts were checked against primary data, actual API responses and visible product behavior.
Core product inference	NVIDIA Nemotron 3 Super through Nebius Token Factory performs bounded tool selection and cited investigation decisions. Nightshift supplies a separate NVIDIA-backed structured review. Application code validates proposals and adds mandatory policy requirements separately.
Pitch deck	Codex with Artifact Session assisted narrative, editable slide authoring, evidence tables and visual QA. The product screenshots are actual captures.
Demo narration	Microsoft Edge neural text-to-speech, generic en-US-AndrewMultilingualNeural voice through the open-source edge-tts client, supplied synthetic narration. No real person's voice was cloned or impersonated.
Captions	Word timings come from the speech engine's word-boundary events. Caption text is the authored narration.
Video and audio production	Claude Code assisted the script, storyboard and editing. Motion graphics were built with Remotion. Product footage is actual Playwright capture of the deployed app, including a live investigation shown at 1× speed; agent-trace values are taken from the committed run receipts. The score and sound effects were synthesized procedurally in Python for the film, and FFmpeg handled loudness mastering.
Visual identity	The original SVG logo and wordmark were authored with Codex assistance, using the project's U/arrow motif. They are vector assets, not a third-party logo or fabricated product screenshot.
YouTube thumbnail	Composed in Remotion from an actual capture of the live investigation result. No generated imagery.

Cursor was not used. AI-generated explanations are not treated as source truth or authorisation. Historical data, model inferences, policy additions, recorded runs and business hypotheses are labelled separately throughout the project.