Headline

Founding AI Engineer | Agentic Systems, LangGraph & LLM Evals | 5 AI backends 0→1 solo | 170K-doc legal RAG serving 2,000+ users

About

Founding AI Engineer at lawworld.ai (Aug 2024 to Jun 2026). Sole AI and backend engineer on a 3-person team, shipped 5 production backends 0→1 (a distributed, microservices-style platform) across 3 architectural generations (LangChain LCEL → LangGraph → 8 agent primitives hand-built in async Python) into an 8-12-engineer-shaped product. I know the internals of the agent frameworks your team runs (LangGraph, CrewAI, AutoGen), not just the API surface: tool registry, prompt assembly, the loop, retry/fallback, state, permissions, observability, eval gate. Ran LangGraph in production 4 months, then built those 8 by hand. 🏗️ What I shipped in production: 1. Reliability — per-page MongoDB checkpointing + Azure Service Bus choreography, at-least-once execution, zero data loss on crashes. 7–8× Pass 1 wall-clock collapse (450–700s → 80–100s) via asyncio.gather + Semaphore(6). 2. Scale — 170K-document legal corpus, 2-stage retrieval (Atlas Vector Search ANN → Voyage rerank-2.5), Perplexity-style citations with a cited-only filter (~30× fewer persisted sources). Serving ~2,000 registered users since the Jan 2026 launch. 3. Security — multi-tenant JWT auth with per-request ownership gates that return 404, not 403, on unauthorized access, closing the enumeration side-channel most teams miss. 4. Cost engineering — per-call cost-attribution middleware into MongoDB. Hundreds of Acts/Rules processed, 100.8M+ tokens, $296.75 total cost, 47.6% cache hit rate. 5. Eval discipline — authored the eval-gating contract (ADR-005): recall@5, phrase-overlap, citation-accuracy thresholds as CI merge gates, written before a line of chat code. ⚙️ Open-source work: ▸ argus — permission-gated multi-agent research engine, 8 primitives built by hand. Clears 7 of 8 metrics on a curated eval benchmark, $0.149/turn unit economics. ▸ bare-agent — published, zero-dependency agent runtime on PyPI. 8 core primitives, bounded ReAct loop. ▸ agents-eval-framework — DSPy Extractor-Judge calibration harness. Kappa 0.12→0.9261 across 5 iterations, 480-trace golden set. 🛠 Core stack: Python 3.12+ (Async) · FastAPI · Pydantic V2 · LiteLLM · MCP (Tool Calling) · DSPy · Docker · Kubernetes + KEDA · Kafka & Message Brokers (Azure Service Bus) · Redis · MongoDB Atlas · PostgreSQL/pgvector · Voyage AI · OpenTelemetry · Terraform · AWS/Azure

Location

Bengaluru

AI stack

antigravity

claude code

codex

langchain

langfuse

langgraph

Work experience

Founding AI Engineer at lawworldai, Aug 2024 - Jun 2026 · 1 yr 11 mos, Bengaluru, Karnataka, India, Architected and deployed deterministic agentic AI systems, focusing entirely on infrastructure reliability, evaluation metrics, and unit economics. 🚀 Agentic Substrate & Orchestration: ▸ Built Primitive-Level Agentic Orchestration: Architected the 8 core agentic primitives from scratch in async Python—including a three-axis loop (tokens, turns, time). Achieved a 7–8× wall-clock speedup (700s → 100s) on Pass 1 extraction via asyncio.gather + Semaphore(6). ▸ Enforced Strict Data Contracts: Designed a multi-agent swarm utilizing production tool-calling with strict Pydantic V2 JSON validation to handle high-entropy reasoning tasks. ▸ Architected Durable Execution: Implemented MongoDB per-turn checkpointing combined with Azure Service Bus choreography to guarantee at-least-once execution and zero data loss on node crashes. 📊 Ops-to-Research Evaluation & FinOps: ▸ Engineered Evaluation Infrastructure: Designed an offline DSPy Extractor-Judge evaluation graph that enforced a 0.90 Cohen's Kappa agreement baseline against human golden datasets. ▸ Deployed Autonomous GEPA Loops: Implemented self-healing pipelines that autonomously rewrite agent instructions based on trace failures. ▸ FinOps Cognitive Routing: Shipped a multi-provider LiteLLM proxy router that dynamically shifts workloads based on cognitive demand. Processed 15,775 pages pushing 100.8M tokens with a 47.6% prompt-cache hit rate, achieving a total inference cost of just $296.75. 🛠 Substrate Stack: Python 3.12+ (Async) · FastAPI · Docker · Kubernetes + KEDA · Terraform · Redis · MongoDB Atlas · DBOS Durable Execution · LiteLLM · DSPy · Pydantic V2 · OpenTelemetry (OTel) · structlog · Model Context Protocol (MCP) · Azure/AWS

Python Developer at The AlgoHype, Jul 2024 - Sep 2024 · 3 mos, New Delhi, Delhi, India, 3-month engagement (pre-lawworld.ai) — built two production AI services: ▸ Synthetic data pipeline for LLM fine-tuning — Azure OpenAI for synthetic Q&A pair generation against a domain corpus. Document ingestion via python-docx + pymongo, Pydantic schema validation + content-relevance + duplicate-detection quality gates, LangFuse for prompt tracing, backoff library for rate-limit retry, Docker for deployment. ▸ Context-aware QA microservice on FastAPI + Docker — multi-provider LLM routing (OpenAI + Groq) via LangChain, Qdrant vector DB for similarity search, Pydantic-typed API contracts, LangFuse instrumentation, Azure Pipelines for CI/CD. 🛠 Stack: Python, FastAPI, Docker, MongoDB, Pydantic, Qdrant, LangChain, LangFuse, OpenAI, Groq, Azure OpenAI, Azure Pipelines.

Junior Machine Learning Engineer at Omdena, Nov 2023 - Jan 2024 · 3 mos, Palo Alto, California, United States, 8-week AI Innovation Challenge with Omdena (community-led, volunteer-basis): collaborated on an LLM-powered interview-prep chatbot — speech input, structured assessment, feedback loop. NLP + RAG + prompt engineering. Demo: https://drive.google.com/file/d/1EDcnxE53pu77E4FJsVt4BMqZIWb5WZMu/view

Data Analyst at Trainity, Apr 2023 - Nov 2023 · 8 mos, Bengaluru, Karnataka, India, 8-month part-time structured training in data analysis + machine learning (concurrent with final-year BTech). Capstone projects: customer-churn prediction model, marketing-campaign optimization analysis, interactive dashboard suite. Stack: Python, SQL, Excel, Tableau.

Education

Bachelor of Technology - BTech, Computer Science Engineering, Parul University

Public links

github.com: https://github.com/subratamondal1/argus

youtube.com: https://youtube.com/@subratamondal1-07