Abhisek Mishra

A production-grade, multi-agent platform automating high-stakes banking Requests for Information (RFIs) for Sanctions Screening, AML, and KYC.

• Orchestrated 6 specialized agents via LangGraph, with 3 MCP servers exposing 15+ secure banking and compliance tools to the LLM — question generation runs on Llama, with a GPT-based reflection agent critiquing and refining output before a human review gate.
• Engineered a safety-first architecture with circuit breakers across 8 external services and automated checkpoint recovery for long-running workflows.
• Integrated Playwright MCP for autonomous UI verification and regression testing of the RFI dashboard.
• Identified and remediated a prompt-injection vulnerability found in security testing — customer-controllable fields were being interpolated directly into the generation prompt — by treating those fields strictly as data and adding input validation.

Private — enterprise production system

A 4-node LangGraph workflow that converts long-form video into short-form clips automatically — deployed and live.

• Transcribe → Identify Clips → Detect Focus → Render pipeline: FastAPI/Python backend using faster-whisper for transcription and FFmpeg for rendering.
• Next.js frontend with real-time progress streaming over SSE.
• GPT-4o-mini identifies the strongest segments from the transcript, combined with face/object detection for smart vertical-crop framing.
• On the roadmap: extending with Temporal for durable execution and checkpointed recovery across long-running renders.

A QLoRA fine-tune of Qwen2.5-1.5B-Instruct that translates Gen Z slang into plain English and back — trained end-to-end on a laptop GPU.

• Rank-16 LoRA across all seven projection matrices, 4-bit NF4 double-quantized, ~20 minutes per training run on an 8GB RTX 4060.
• Diagnosed why a correct answer present in the training data still wasn't learned — LoRA only updates ~1.2% of parameters, so a couple of examples can't override a strong pretrained prior.
• Audited my own 'grounded in real datasets' claim against the code, found the pipeline never read the source files, and rebuilt it to generate ~4,600 examples programmatically.

A hands-on field guide to RAG, built inside a working app — every retrieval and evaluation technique implemented as a switchable strategy rather than a slide.

• Built rag-ollama-js to implement and compare RAG strategies against the same document set inside one codebase, instead of reading about them.
• 46+ GitHub stars and 15 forks — public signal from other engineers building similar systems.
• Spun out the evaluation layer into raglens, a RAGAS-style RAG evaluation library for TypeScript, published to npm (provider-agnostic, zero required runtime dependencies).

A bank-wide Claude Code plugin standardizing legacy service migrations with deterministic, review-consistent output.

• Architected a bank-wide Claude Code plugin — CLAUDE.md context standards, reusable Skills, subagents, and hooks — to standardize and structurally enforce Java-to-Node.js legacy service migrations across the engineering organization.
• Built a Git hook to measure Claude-authored vs. manually-written code volume per engineer, feeding into adoption and cadence reporting.
• Estimated ~40% reduction in migration time based on comparative task estimates; adopted bank-wide across the engineering organization.

Private — enterprise production system

A highly scalable, event-driven microservices platform underpinning Tablet Banking, ATM servicing, and CRM integration.

• Node.js for high-throughput APIs, MongoDB for flexible data modeling, Redis for distributed caching.
• Applied Domain-Driven Design and service resiliency patterns to decouple core business capabilities.
• Owned the full solution lifecycle across multiple modules — design, deployment, monitoring, and post-launch support.

Private — enterprise production system