KAUSHIK
REDDY.

An agent can read back the right order and still submit the wrong one. I built IndicOrderBench to check what actually reaches the backend.

The build

I connected scripted, multi-turn English and Hinglish conversations to an isolated ordering backend. Python and HTTP adapters give agents the same tools. Reports keep the conversation, tool calls, and committed order together, while a CI gate catches regressions.

The decision

The verifier checks acceptable order states deterministically, without an LLM judge. Infrastructure failures and caller interactions the script cannot handle are reported separately from genuine ordering failures.

What it shows

The buggy-versus-fixed demo catches quantity and modifier errors hidden behind a convincing readback. It demonstrates the harness rather than a production voice stack; the evaluation is turn-based, and Hinglish scenarios still need native-speaker review.

I built Thrill AI around Rasa, a multilingual table assistant connecting a scanned QR code, spoken or typed requests, a live cart, and confirmed kitchen rounds.

The build

Next.js, React, PostgreSQL, and Sarvam AI connect the guest experience to the restaurant dashboard. Guests can order in Indian languages, review each round, and request the bill; owners manage menus, live orders, and table sessions.

The decision

The model suggests structured operations against real menu IDs. The server validates those operations, checks availability, and calculates prices in paise. A kitchen round requires explicit guest confirmation; transaction locking and idempotency keys handle repeated submissions without duplicating the order.

Where it stands

The repository records a local owner-to-guest walkthrough through two kitchen rounds and settlement. Hosted rollout, live provider calls, and payment verification still need separate validation. The Rust services remain experimental and are not connected to the main app.

Deleting a record leaves a harder question: what happens to its encrypted copies? I built Aegis around per-subject keys and an explicit erasure workflow.

The build

The current aegis-shred toolkit combines a Rust vault, Python bindings, and a CLI. It encrypts records and files under per-subject keys, stores wrapped keys in SQLite, and records receipts and tombstones when a key is erased.

The decision

Vault operations consult the keystore rather than a key cache, so processes sharing it see an erasure on their next call. File decryption writes to a temporary destination and publishes the output only after all chunks verify.

The boundaries

An erasure journal lets applications reapply deletions after restoring an older keystore. Plaintext copied outside the vault remains outside that protection. I keep the format, threat model, and operational limits documented alongside the code.