funded-report-assistant

Tracker lint, source-referenced progress reports and review packs for a funded R&D project. By Pavan Yadav Annappa. Code on GitHub

All project data on these pages is fictional. "eDrive-X" and Partners A-D are generated by a seeded script; no real company, person or project data is used.

Overview

A small command line tool (fra) for the paperwork side of a publicly funded R&D project: it checks an Excel project tracker against the meeting minutes, drafts the periodic progress report with a source reference on every sentence and table row, and prepares the agenda and open-points list for a milestone review. An optional LLM step rewrites the work-package summaries as prose, behind a guard that rejects any draft with a number, date or ID that is not in the source rows.

Live demo: https://huggingface.co/spaces/pavanyadava07/funded-report-assistant (sample tracker, lint findings, generated report, LLM variant, review pack, evaluation)

All project data is fictional. "eDrive-X" (a GaN eDrive inverter and eAxle demonstrator) and Partners A-D are produced by a seeded generator (fra/generator.py). No real company, person, funding call or project data is used.

What it does

data/edrive-x/tracker.xlsx WorkPackages, Milestones, ActionItems, Risks (+ Meta) data/edrive-x/minutes/*.md 7 meetings: action status, decisions, new action items -> fra lint 13 consistency rules, findings as Markdown + JSON -> fra report --period 2026-H1 progress report (Markdown + .docx), every line cites sheet!row or minutes file:line -> fra report --llm ollama:... same report, WP summaries reworded by an LLM if they pass the guard -> fra review-pack --milestone M3 agenda, open points, earlier decisions, action log extract

Lint rules

Measured results

The block below is written by scripts/make_results.py from results/*.json; numbers are not edited by hand. Details per rule and rejected-sentence examples: docs/RESULTS.md.

Tracker lint on 20 generated trackers (seeds 1000-1019, 13 defect kinds planted per tracker where placeable, formatting noise on) with 257 planted inconsistencies, plus the same 20 trackers without defects:

Mean runtime per tracker on this host: lint 0.3 ms, deterministic report 0.7 ms (after loading the .xlsx).

Closed world: the generator and the checker were written together by the same author, so 100 % recall here shows that each rule fires on the defect it was written for and stays quiet on clean data. It does not show how the rules do on a real project tracker. The baseline false alarms are cancelled actions with a past due date, which a plain "not closed" filter reports as overdue.

LLM narrative + guard: llama3.1:8b via Ollama 0.33.3, temperature 0, seed 42, device: CPU only (num_gpu=0, 32 vCPU). 21 projects (the sample + 20 generated), period 2026-H1, 126 work-package summaries.

Rejected summaries fall back to the deterministic text, so the report never contains an ungrounded number. The guard checks tokens, not meaning: a sentence can use only true numbers and still bind them to the wrong claim. In this run the status check found no contradiction, yet manual reading of the kept summaries found 'ahead of plan' for late milestones and 'exceeds the plan by 153%' for 153 % of plan. The other way round, several rejected sentences hold a correct derived number (for example 'exceeding the plan by 8%' for 32.5 vs 30.0 PM): the guard rejects any number that is not literally in the facts. The LLM ran on CPU because the host GPU (NVIDIA L4) was in use by another job; runtimes are CPU runtimes. No time saving is claimed; only runtime was measured.

Pre-rendered by scripts/build_site.py from the files in the repository.