TallyBard

Open-source collections copilot for people who sell on credit. Freelancers, tutors, shops running a baki/udhaar book, tailors, caterers, small wholesalers, community treasurers — anyone who delivers first and gets paid later.

TallyBard imports invoices and informal credit entries, predicts which open receivables will be paid more than 7 days late, ranks a chase queue by amount × risk, and drafts polite WhatsApp nudges in the client's language (Bangla, English, or mixed) at friendly → firm → final tones. A human approves every message — nothing is ever sent automatically.

Demo data is synthetic. All clients, receivables, and notes in the demo dataset are generated (db/generate_synthetic.py) and labelled synthetic in the UI and docs.

Built with PriorLabs-TabPFN.

Try it

• Zero-login demo (/app): opens directly onto the synthetic ledger (60 clients, 600 receivables, Bangla + English notes). No sign-up, no tokens, no cookies.
• Sample CSV: download /sample-ledger.csv or click Sample CSV on /app/import to test extraction. Confirmed rows stay in your browser (localStorage) in demo mode and never overwrite the shared synthetic ledger; click Reset demo in the banner at any time.
• Full frontend & Render guide: see docs/FRONTEND.md and screenshots in docs/screenshots/.

Architecture

flowchart LR
    U[User] --> W[Static web UI]
    W -->|HTTPS, CORS-locked| A[Mastra API — TypeScript]
    A -->|anonymised numeric features only| S[Risk sidecar — FastAPI + TabPFN]
    A -->|prompts| L{Model backend}
    L -->|self-hosted| O[Ollama on DO droplet: gemma4 + embeddinggemma]
    L -->|fallback| G[Google AI Studio: gemma-4 + gemini-embedding]
    A --> T[(Tiger Cloud Postgres)]
    T --> HT[payment_events hypertable]
    T --> HS[hybrid search: pg_textsearch BM25 + vectorscale DiskANN + RRF]
    A --> SE[Sentry Agent Tracing — genAI inputs/outputs OFF in prod]
    S --> T

Loading

Deploy: render.yaml blueprint — static site + Node API + Python sidecar.

Quickstart

# 1. database (Tiger Cloud Postgres 17+)
uv run scripts/db.py -f db/schema.sql     # extensions, hypertable, RRF fns
uv run db/generate_synthetic.py           # writes db/data/ledger.csv
uv run scripts/db.py -f db/seeds.sql      # seeds synthetic demo ledger

# 2. risk sidecar (Python 3.12, tabpfn local weights)
cd apps/risk-sidecar && uv venv --python 3.12 .venv && uv pip install -r requirements.txt
.venv/Scripts/python -m uvicorn main:app --port 8000     # .venv/bin/python on unix

# 3. Mastra API
cd ../api && npm ci && npx mastra dev

# 4. web site + demo app (zero external deps, deterministic build → dist/)
cd ../web && npm run check && npm run dev # → http://localhost:5173

Copy .env.example → .env and fill in the values (never commit .env).

The core loop

1. Import — paste CSV or informal text; Gemma extracts rows; you confirm before anything is written.
2. Score — POST /score on the sidecar refits TabPFN on the labelled history and returns P(paid >7d late) per open receivable with a SQL-derived "why" (amount, days overdue, client's prior late rate — real stats, not model explanations).
3. Queue — open receivables ranked by amount × p_late7.
4. Draft — nudge in the client's language and chosen tone, grounded by hybrid BM25+vector search over that client's notes and your previous nudges so wording never repeats.
5. Approve — the API marks the draft approved and returns a wa.me click-to-send link. You press send. Then mark_sent, and mark_paid when money lands.
6. Weekly run — a Mastra workflow scores + drafts the top of the queue and suspends until you resume with approved: true.

Partner tech

Piece	Used for
Tiger Data (Postgres 18 + TimescaleDB)	ledger, payment_events hypertable, BM25 (pg_textsearch) + DiskANN (pgvectorscale) hybrid search fused with RRF
Prior Labs TabPFN (3.5 local + v2 comparison)	P(paid >7d late) on tiny ledgers — honest head-to-head vs a past-late-rate baseline
Gemma (gemma4 via Ollama; gemma-4-26b-a4b-it via AI Studio)	text import extraction, nudge drafting, ledger chat
EmbeddingGemma / gemini-embedding-001 @768d	vector leg of hybrid search, nudge dedup
Mastra	tools, agent, suspend/resume weekly-chase workflow, Postgres memory
Sentry Agent Tracing	instrument.mjs + @mastra/otel-bridge; genAI inputs/outputs off in prod
DigitalOcean	Ollama droplet (CPU demo droplet + optional 3h GPU benchmark droplet)
Render	blueprint deploy of all three services
GitHub Copilot	starter PR + issues #2/#3 assigned to the coding agent; docs/copilot-cli-task.md is the Entire-captured CLI task
Entire	session capture for the Copilot CLI task
DevRelay	challenge/MLH wiring for this submission

Results (real runs — see docs/results.md)

Synthetic benchmark, time-split, tabpfn==9.2.0 local CPU, 2026-10-16:

train N	baseline AUC	TabPFN-3.5 AUC	TabPFN v2 AUC
50	0.655	0.533	0.515
100	0.653	0.619	0.625
300	0.670	0.672	0.681

The free baseline wins at N=50 and stays competitive — TallyBard shows both numbers instead of pretending the fancy model is always right. TabPFN-3.5 weights are non-commercial (demo/benchmark only); v2 is the permissive Prior Labs licence. docs/results.md has the full table, commands, and dates.

Honest limitations

• Synthetic data only — nothing here is evidence about real clients.
• /score refits TabPFN on every call (~1 min cold on CPU, less warm).
• TabPFN-3.5 weights are non-commercial; production/self-host use needs v2 weights (permissive) or your own model. Verify licence terms yourself.
• gemma-4-31b-it hit AI Studio free-tier rate limits during testing — the fallback default is gemma-4-26b-a4b-it.
• No auth/multi-tenancy, no auto-send, no payments — by design.
• Public-dataset backtest was skipped: no payments dataset with a clearly verified licence was found in a 20-minute search.

Prior art

PayMind (by fracabu) — a previous open-source AI-collections project — inspired the space. TallyBard shares no code with it; it differs in being self-hostable, model-swappable, handling informal credit (baki/udhaar), scoring with TabPFN against an explicit baseline, and hybrid-search-grounded language-aware drafting.

Repo layout

apps/web/          marketing site + zero-login demo SPA (build.mjs → dist/, 20 node:test checks)
apps/api/          Mastra API — tools, agent, weekly-chase workflow, Sentry
apps/risk-sidecar/ FastAPI + TabPFN scoring, point-in-time features, backtest
db/                schema.sql, seeds.sql, synthetic generator
scripts/           db.py helper, bootstrap-ollama-do.sh
docs/              PLAN, STATUS, FRONTEND, results, backtest.json, screenshots/
render.yaml        Render blueprint

Post-deadline commits

(none yet — every commit before 2026-10-18 23:59 PDT)

Licence

MIT — see LICENSE. Model weights keep their own licences (Gemma Terms of Use; TabPFN per Prior Labs — see results.md).