HifazatTabحفاظت · protection in the weights

A Pakistani kiryana shop agent over 200+ SKUs — price, stock, and loan eligibility in Urdu, Hindi, and English. Tricks refused in every script.

Live demo = same shop tool contract (instant, no GPU). Catalog: demo — dashboard · about.

Pick a chip, type a question, or use the mic — answers stream into this thread.

اردو mic → Groq Whisper. EN/HI → browser STT.

Why Track C

Lexsi Agentic Fine-tuning: train with AgentTune, prove with AuditKIT. Not a chatbot wrapper — tool reliability learned in weights.

  1. Closed shop CSV — price, stock, details, category, low stock, loan tiers in PKR.
  2. Tiny Aya fire — South Asia multilingual base. No native tools template → tools_fallback_prompt.
  3. AgentTune GRPO — verifiable rewards for exact tool+arg; inventing tools is punished.
  4. AuditKIT — before/after on trajectories.jsonl (validity, F1, task success).
  5. Script-switch safety — same attack in Urdu ↔ Hindi ↔ English must stay refused after quant.

Six tools

Closed shop contract — answer only through these. Everything else is refused.

lookup_price

PKR price + stock for a catalog item.
arg → item key

check_stock

Stock only; flags low inventory under 20.
arg → item key

item_details

Full SKU card: price, stock, category.
arg → item key

list_category

Browse an aisle (staples, oils, dairy…).
arg → category

list_low_stock

SKUs at or below threshold (default 20).
arg → threshold

loan_eligibility

Tier from monthly income. Never approves.
arg → income

Before → after

Zero-shot Tiny Aya learned JSON shape but failed content/safety. Gold-JSON SFT closed the gap. Live demo = rule agent; metrics adapter: hifazat-tab-sft-v1.

Before · zero-shot

tool validity98%
task success11%
refusal precision14%

After · SFT LoRA

tool validity100%
task success98%
refusal precision88%

AFTER SFT (801 tool / 24 probe): 801/801 · 784/801 · 21/24. Targets ≥85% / ≥75% / ≥80% — met. GRPO v1 was flat; SFT wins Track C AFTER.

60-second demo

What judges should see without explaining the stack first.

  1. English sample → price of atta → hear 2850 PKR.
  2. Switch to اردو → same item → same number, Urdu script UI.
  3. Fake loan-approval sample → spoken refusal (no tool leak).
  4. Run red-team → 24/24 blocked across scripts, incl. roman-Urdu.
  5. Point to HF adapter + BEFORE gap (11% task → target 75%+).

Safety that ships

Refusals cover fake approvals, PII fishing, forced loan advice, prompt injection, and demanded lies — in English, Devanagari, and Nastaliq.

Not run yet.

Shop oracle · … SKUs

The only numbers the agent may speak. Ask any listed item by English, Urdu, or Hindi name.

Artifacts

Everything public as umeraamir69 for Lexsi credit match.

HifazatTab · Lexsi AI Safety Hackathon 2026 · Track C Agentic Fine-tuning · AgentTune → AuditKIT