I build systems that score, rank and verify — then measure whether they worked.
Data analyst and systems builder · Founder of Thrill AI · MS Data Analytics, George Mason '25
Two installable packages, both at v0.1 with CI and a no-keys demo:
| Package | What it does | Install |
|---|---|---|
| IndicOrderBench | Benchmark and CI gate that catches voice ordering agents committing the wrong order, in English and Hinglish. It checks the order the agent submitted to the backend, not the transcript, so a convincing read-back can still fail |  |
| aegis-shred | Crypto-shredding for application data: one encryption key per user, so erasing a person destroys one key and every copy of their data becomes unreadable, backups included. Rust core, Python package, aegis CLI. Alpha, not yet independently audited |
 |
A multi-agent village strategy game, and an inspectable AI simulation lab. Set one objective; six residents bid for jobs, walk to workplaces, gather, build and recruit. You can open any resident and read why it took that job, what it remembers, and what it spent. Runs locally with no API key, account or database.
| Inside it | |
|---|---|
| Agents that explain themselves | Jobs are claimed by bidding on role fit, distance and bounded experience; every decision, memory and spend is inspectable |
| A real game around it | Six building types, two builders, treasury, knights/archers/catapults, three enemy strongholds |
| A research desk | Matched seeds, replayable evidence, JSON/CSV export — every recorded result reproducible |
| Optional live models | Claude can propose a social action in a local-worker experiment, with validated actions and spend reservations |
Fixes and features shipped into projects I use. Every bug here was reproduced before it was fixed, and every fix was verified through the path the docs describe:
| Project | Contribution | Status |
|---|---|---|
| libredb-studio ★1.0k | One-command Keycloak SSO demo stack (Compose + Caddy TLS + preconfigured realm), two follow-ups binding the demo stacks to loopback only (#960, #1004) and three security-docs fixes (#938, #939, #940) | ✅ 6 merged · 🔄 #1119 in review |
| BizzAI ★30 | Return-refund options with a customer-credit system (#186) and an infrastructure upgrade (#178) for this open-source POS and inventory system | ✅ merged |
| Backlog.md ★6.9k | Git fetch no longer fails when the only remote isn't named origin, with a regression test |
🔄 in review |
| PasarGuard ★2.6k | Clipboard copy inside focus-trapped menus on plain-HTTP deployments | 🔄 in review |
| tunarr ★2.6k | Fixer for mislabelled Jellyfin/Emby rows that made guide endpoints return 500, with idempotency and false-positive guard tests | 🔄 in review |
| NVIDIA NeMo Curator ★1.8k | Batched processing in the MinHash dedup stage | 🔄 in review |
| httpx2 ★1.5k | Contributing docs point at the discussion categories that exist | 🔄 in review |
| Project | What it does | Stack |
|---|---|---|
| Thrill AI 🏢 | The company I founded: AI voice ordering for Indian restaurants, taking orders in 22+ languages and automating the service loop (product site; source is private) | Voice AI · Multilingual NLU |
| aegiseval | Adversarial safety evaluation for a tool-using RAG agent: threat model, 260-prompt golden set, validated LLM-as-judge, mitigations measured before and after | Python · RAG · red teaming |
| MerchantLens | Merchant analytics lakehouse on Databricks where access rules are enforced by the query engine, not by asking the agent nicely | Databricks · Delta · Unity Catalog |
| pulse | Realtime task board that shows its own latency budget live: <100 ms per interaction, <200 ms to sync | Next.js · tRPC · realtime |
| Contract & Invoice Intelligence | Checks invoices against their contracts, flags mismatches and risky clauses, tracks precision/recall/F1 | LangGraph · Qdrant · Next.js |
| rocm-devops-starter | Probe, smoke-test training run, ROCm container and CI for PyTorch on AMD GPUs, plus a portability scanner that reads a repo without importing it. Status table says exactly what has run on real hardware | Python · PyTorch · Docker |
Analysis
AI systems
Engineering
Ops
Taking IndicOrderBench and aegis-shred from v0.1 to their first outside users · Prometheus, Terraform and Kubernetes · more agent work in Settlement · open to data analyst, analytics engineering and AI engineering roles.
If a project here is useful to you, a ⭐ or an issue is always welcome.