Ankit Pradhan

pnpm

Contributor

Fixed an auth config bug affecting both the TypeScript CLI and the newer Rust port. It meant understanding two parallel implementations of the same behavior and reasoning about cross-platform config resolution.

View Pull Request →

@bro-ankit/nestjs-typeorm-transactional-context

Author • Published npm Package

A NestJS library for transactional context propagation across async boundaries using AsyncLocalStorage. No decorators, and it works outside the HTTP request lifecycle. 80+ weekly downloads, used in production fintech systems.

View on NPM →

nestjs-sqs

Contributor

Fixed a lifecycle bug in message handler wrapping. Tracing it meant digging into NestJS's internal dependency injection and module lifecycle, not just the application code on top of it.

View Pull Request →

AI Support Triage Copilot

AI Engineering • Streaming • Fine-Tuning • LangGraph

A copilot that triages support tickets: intake, investigation, and a proposed action (refund, credit, escalation) that stops at a human approval gate before anything executes. I built it to find where an LLM pipeline like this actually breaks, then fixed those places one at a time.

What I tried, what broke, what I measured

• Fine-tuned a Qwen2.5-0.5B classifier locally (MLX LoRA, served via Ollama) and benchmarked it against Gemini on this project's own labeled set. Result: 96.4% accuracy versus 57.1%, about 4x faster, and free to run. Hybrid routing escalates to Gemini only on a schema-validation failure.
• The investigation pipeline (classify → recall → retrieve → diagnose → propose) runs as a LangGraph StateGraph, streamed stage-by-stage over SSE to a dashboard. I tested the abort path specifically: disconnect mid-classify, confirm recall never runs.
• I ran a prompt-injection attack against the live Gemini API. It worked, and hijacked a proposed refund. Fixed it with a structured systemInstruction boundary, delimited untrusted content, and canary-token leak detection, then re-ran the same attack to confirm it no longer lands.
• Every diagnosis has to cite the specific knowledge-base excerpt backing it, or it routes to a human instead of completing silently. Retrieval escalates to a second, query-decomposed hop only when a cross-encoder confidence signal says the first pass was weak.
• Auth is an OAuth 2.1 resource server (Auth0-issued RS256 JWTs) gating every route by scope, with tenant isolation running through the retrieval layer itself. The mutating action needs a separate, race-tested approval compare-and-swap from the token that proposed it.

View Source → View Architecture Diagram →

AI Incident Response Copilot

AI Engineering • Multi-Agent Orchestration • NestJS

A multi-agent system that investigates a production incident and proposes a ranked root-cause hypothesis with remediation steps. Every code path stops at propose: nothing executes a fix, a person reads the output and decides.

How the agents actually behave

• A Log Analysis agent and a Runbook Search agent each run multi-turn tool-calling loops against stdio MCP servers: search, read, decide whether to search again.
• Remediation weighs proposed fixes against a Neo4j service-dependency graph, queried through a plain repository call rather than an MCP tool, since the model never has to decide which query to run there.
• The two investigative agents dispatch via Promise.allSettled, so one failing doesn't cancel the other. Every stage is nullable, and a failure surfaces as a plain-English warning rather than a thrown 500.
• Built an LLM-as-judge eval harness and ran it against 6 seeded incidents. It scored 0.96 average correctness and 1.00 average groundedness, and the whole run cost $0.03.

View Source → View Architecture Diagram →

Smart Semantic Bookmarking & Memory Engine

AI Engineering • RAG Pipeline • NestJS

Turns saved bookmarks into queryable semantic memory, built with the same review discipline as the backend services I ship at work.

What I checked before calling it done

• Vector search (pgvector HNSW) and Postgres full-text search get fused via Reciprocal Rank Fusion, then reranked by a locally-run cross-encoder. No added API cost per query.
• AI-generated summaries and tags sit in a review queue before they're ever embedded. Corrections get logged as a fine-tuning signal for later.
• Ran an LLM-as-judge harness against a 17-case golden set. All 17 passed, averaging 0.94 relevance and 1.0 faithfulness.
• Gemini's native function calling drives a multi-step loop that searches, synthesizes, and creates follow-up todos on its own.

View Source → View Architecture Diagram →

Campus Placement Pipeline

Data Engineering • Databricks • Machine Learning

A Databricks data engineering project simulating a campus placement analytics system end to end, from raw ingestion to ML-served predictions.

What's in it

• Bronze → Silver → Gold Delta Lake pipeline with per-layer schema validation and a dead-letter quarantine for bad records.
• Simulates 100,000 students across 150 universities, 500 companies, 50,000 applications, and 80,000 interviews via batch and streaming ingestion.
• A Random Forest classifier trained on enriched student features predicts placement outcomes.
• A FastAPI layer exposes live predictions and pipeline statistics through Databricks Connect.

View Source →

toGames

Algorithms • Canvas Rendering • Vanilla JavaScript

Two textbook algorithms, built into things you can actually play, in plain JavaScript with no framework and no build step. No résumé reason, I just wanted to build them.

• Unbeatable tic-tac-toe: recursive minimax evaluates the full game tree and returns the optimal move every turn.
• Maze generation you can watch happen: randomized depth-first search with an explicit backtracking stack, carved live on canvas.
• Selectable grid size, arrow-key navigation, and a one-click PNG export of the generated maze.

▶ Play Live → View Source →