Make correctness checkable
An LLM's output is a guess until something can verify it. Line-item reconciliation, schema validation and eval harnesses turn guesses into claims you can gate on.
I'm Nakshatra Joshi — a software engineer working where LLMs meet real backends. Typed services on PostgreSQL, Redis-backed job pipelines, and tool-calling agents that are gated by checks rather than trusted on vibes.
Each one is written up as a case study — the constraint, the decisions, and what I'd do differently.
Every one of these exists because a system taught it to me the expensive way.
An LLM's output is a guess until something can verify it. Line-item reconciliation, schema validation and eval harnesses turn guesses into claims you can gate on.
Agents get a tool per legal transition, not free rein over the workflow. The model decides what happens next; the system decides what is allowed.
Upstream providers rate-limit and fall over. Typed error paths, retries with sweep intervals, and fallbacks keep a session alive instead of dropping it.
Type-check and lint gates, SHA-tagged images, health-checked non-root containers. A rollback should be redeploying a known-good tag, not an incident.
If I can't explain how an agent actually works, I don't understand it well enough to ship one.
Three roles, two of which put code in front of paying customers.

May 2026 — Aug 2026
Pune, Maharashtra · Hybrid
Two production systems: a chat-based freight-procurement agent and the backend of an AI document-extraction platform.

Jun 2025 — Aug 2025
Delhi · Remote
RAG content-generation pipelines with human-in-the-loop approval, and the MLOps plumbing to swap models without redeploying.

Jul 2024
On-site
Dashboarding for sales analysis, and the data pipelines feeding it.
Listed by what I've shipped with, not what I've read about.
Available Jan 2027 — 6-month internship or full-time