kenmandal/ERP-DocIQ

ERP-DocIQ — back-office automation on small models

Build Small Hackathon · Practical track (Backyard AI). A useful, problem-solving app that runs on hardware you own — built entirely on open models ≤32B, with OpenBMB MiniCPM as the load-bearing model (and fine-tuned for the domain).

💡 The idea

Retail back-offices drown in paperwork and report requests. ERP-DocIQ is an open-source, UiPath-style assistant that reads your documents, answers questions about your ERP in plain English, and automates the boring clicks — all on small models, no per-robot license, no data leaving your tenancy.

🧩 The business problem

A retail brand's IT team runs UiPath for invoice/PO processing and portal automation. It's expensive (per-robot / AI-Unit licensing, six figures/yr at volume), brittle (selector recorders break on every UI change), locked-in (closed IDP models you can't swap or fine-tune), and slow to change (new layouts wait on a closed retraining toolchain). On top of that, every "what did we spend in Q2?" becomes a BI ticket. They wanted an AI-native, open, cheaper alternative they control.

🛠️ The solution approach — one Gradio app

1. 1.Read any document (OCR + IDP). A hybrid pipeline — OCR → classify → extract → normalize → enrich (RAG) → validate → post / human-review — reads orders, receipts, invoices, contracts and complex multi-layer forms, even messy scans/photos, into structured ERP records.
2. 2.Ask your ERP reports (ERP DocIQ chatbot). Natural-language NLQ → SQL, analytics, summaries and "why" reasoning over a simulated retail ERP (vendors · POs · invoices · GL · inventory · returns). Every figure comes from real SQL over the data — the model only phrases the answer, it never invents numbers.
3. 3.Automate the clicks (agentic browser). A self-correcting, multi-step browser agent drives a portal (dashboard → Procurement → Create Order → read the complex form) — selector-free and self-healing, replacing fragile RPA recorders.

🤖 Models used — small (≤32B), the right one for each job

Three labs, eight models, all under the cap. MiniCPM is the core.

A single ~8B MiniCPM-V powers OCR and vision extraction and the grounded chat phrasing — so the whole product runs on small, swappable, open weights. GET /api/models lists which are live.

🎯 Fine-tuning — adapting a small model to the ERP domain

We fine-tune OpenBMB MiniCPM3-4B (4B) on an instruction dataset built from the ERP knowledgebase (results/erp_sft.jsonl):

• —Production: a LoRA (PEFT/TRL SFTTrainer) recipe — scripts/finetune_erp.py --backend hf.
• —Offline CPU demo (runs anywhere): trains the ERP NLQ-routing head on the same data with a real train/test split — untrained 8.3% → fine-tuned 91.7% (+83 pts), 100% routed-SQL execution. See results/erp_finetune_report.json.

📈 Benefits / value delivered

• —Significantly lower inference cost vs per-document RPA licensing (model routing + prompt caching).
• —No vendor lock-in, data stays on-prem — open weights you can run, read and fine-tune.
• —Self-service ERP analytics (NLQ) deflects routine BI/report tickets off the queue.
• —Reads documents classic OCR can't — MiniCPM-V cuts character error ~5.6× vs Tesseract (CER 2.6% vs 14.7%; see results/ocr_quality_report.json); field-exact 90.7%.
• —Honest, measured — every claim is backed by a published eval/quality/fine-tune report.

✅ How it meets the Build Small criteria

• —≤32B params — every model is ≤32B; the reasoning/NLQ engine + fine-tune target is 4B.
• —Best MiniCPM Build — MiniCPM-V (OCR + vision) and MiniCPM3-4B (reasoning/NLQ + fine-tuned) are the core of the experience; vision/omni variants qualify.
• —Best Agent — multi-step, self-healing agentic browser automation + an IDP state graph.
• —Ships as a Gradio Space in the Build Small org; runs offline (deterministic ERP engine + sidecar OCR) and upgrades to hosted MiniCPM when keys are set.

▶️ Use it (tabs)

• —Process a document — pick an OCR backend (auto, minicpm, tesseract) + a sample (try extreme_receipt_photo or complex_invoice_messy) or upload your own → multi-layer fields + KPIs.
• —ERP DocIQ (chat) — ask "Why did spend rise in Q2 2026?", "Top vendors by spend", "late-payment rate" → grounded answer + SQL + the fine-tuning panel.
• —Search (RAG) — semantic vendor-master retrieval. Web Automation — multi-step browser flow.

📦 Published results (results/)

• —ocr_quality_report.json — OCR CER/WER + field accuracy per backend.
• —erp_finetune_report.json + erp_sft.jsonl — fine-tune metrics + the instruction dataset.

⚙️ Configure (Space → Settings → Variables and secrets)

• —Variables: MINICPM_BASE_URL=https://api.modelbest.cn/v1, MINICPM_MODEL=MiniCPM-V-4.6-Instruct
• —Secret: MINICPM_API_KEY=… (OpenBMB/ModelBest). Tesseract ships via packages.txt.
• —Runs fully offline without a key — ERP DocIQ uses its deterministic SQL engine and OCR falls back to the sidecar, so every tab works.

🎥 Demo & social

• —Demo video: https://youtu.be/mWs7eRVH_GM
• —Social post: https://www.linkedin.com/posts/kaniskamandal_huggingface-buildsmall-buildsmall-share-7472163579094401024-fjp9/
• —Git repo: https://github.com/agency-world/Project-Aperture
• —Blog post: https://huggingface.co/spaces/build-small-hackathon/ERP-DocIQ/blob/main/BLOG.md