I build systems that can be checked, not trusted.
LLM tooling on top of production data infrastructure — typed plans, governed semantic layers,
and an audit trail behind every number.
Final-year B.Tech CS (Computer Science & Business Systems) at SASTRA University, graduating June 2027. Most of my time goes into one question: when an AI system hands you a number, how do you know it's right? My answer so far is to take the freedom away from the model — make it emit a typed plan, compile that plan deterministically, and ship the evidence alongside the answer.
Off-hours I send patches to the data tools I use.
|
Receipts · payments analytics agent Built for the Razorpay AI Builders programme. Ask a question in English, Tamil or Hindi → get a number plus a receipt: the plan, the generated SQL, and the rows it came from. The model never writes SQL. It emits a typed Guarded by an integrity charter enforced in CI: an eval suite plus fault-injection and meta-tests that check the guardrails actually fire. |
Walmart Data Platform · CDC + quality gates Syncs an operational Postgres into a Databricks lakehouse hourly via change data capture — only rows inserted, updated or deleted since the last run move across six related tables. Ingestion path chosen per source by how fast it changes: transactional tables follow the change stream, low-velocity reference data is read in place from S3 as an external location — no copied duplicate that can drift. dbt tests at the conformed layer (uniqueness, not-null, referential integrity, accepted values) fail the run instead of letting bad rows reach a dashboard. Airflow in Docker owns scheduling, retries and alerts. |
|
SpecGuard · spec-drift detection Does the code still do what the spec says? Evidence-cited verdicts, an adversarial second pass, and abstention instead of guessing. |
MCP Server Bridge · secure DB access over MCP A pure MCP data-access layer: sandboxed SQL execution and schema introspection, with NL→SQL left to the client. |
| apache/airflow#72625 | region_name was silently ignored by the Step Functions execution trigger — deferred tasks polled the wrong AWS region. Merged, shipped in apache-airflow-providers-amazon 9.36.0. |
| PrefectHQ/fastmcp#5030 | A tool with an output schema that returns an unserializable value reports success — handing the client a repr() string and no structured content. Filed with a reproduction and a fix branch; more findings queued. |
Titan Company Limited (A TATA Enterprise) — Live Project Intern, Systems Dept · Dec 2025 Integrated a hosted image-generation model into a customer-facing virtual try-on flow over its REST API — shipped end to end in three weeks.
🏆 Meta PyTorch OpenEnv Hackathon × Scaler School of Technology — final round, top 100 teams · Apr 2026
Languages: Python, SQL · AI & Data: LLM agents, semantic layers, dbt, Airflow, CDC, Delta Lake, data quality testing · Tools: Databricks, AWS S3, Docker, Git/GitHub, GitHub Actions
📍 Hosur, Tamil Nadu · 📬 nandeshjeyalakshmi@gmail.com