Lex Legal Chatbot
Framework-free ReAct loop routing across 3 retrieval tools, grounding every fact back to its original source PDF across 8 Indian languages.
170,439Total Documents Indexed
787 Acts · 43,927 Judgments · 125,725 Headnotes
Résumé
Subrata Mondal
Founding AI Engineer | AI Infrastructure & Agentic Systems Engineer
Bangalore, India · +91 6294505807 · subratasubha2@gmail.com
01 — Professional summary
Founding AI Engineer who took an AI legal tech product from ideation to production as the Solo AI Engineer, serving 2,000+ users. Built high-throughput agentic systems in async Python, saved $12K/year replacing a paid AI-native rich-text editor, cut LLM costs 5-10x, and dropped P99 latency to 50ms. Catches LLM regressions before code ships to production via strict CI/CD evaluations.
02 — Technical skills
Backend
Python, FastAPI, Microservices, PostgreSQL, Vector DB, PG Vector, MongoDB Atlas, Redis, Message Queue Broker, REST APIs, WebSockets, SSE, AsyncIO, OpenTelemetry
Gen AI
LangChain, LangGraph, Multi Agent Orchestration, MCP, DSPy, LiteLLM, AI Agents, Agentic Workflows, RAG, CRAG, Hybrid RAG, Prompt Engineering, Evaluations (RAGAS, LLM-as-a-Judge), Hallucination Reduction, LangFuse, Guardrails
Infra
Docker, Kubernetes, AWS, Azure, CI/CD, KEDA, System Design, Latency Tuning, LLM FinOps & Cost Optimization
Tools
Claude Code, Cursor, Git, UV, Ruff, PyTest
Frontend
Next.js, React, TypeScript, Tailwind CSS, shadcn/ui, Zustand, React Query, Framer Motion
03 — Professional experience
Founding AI Engineer @ LawWorld
Aug 2024 – Jun 2026Bangalore, India (On-site)
Architected an AI native legal tech product end-to-end — production AI applications and scalable backend systems. Designed agentic AI workflows and RAG pipelines serving 2,000+ users over a 170K-document corpus, prioritizing reliability, latency, and operational efficiency.
In-house editor replacement: Replaced a $12K/year vendor subscription in-house, solo-building the agentic rich-text editor backend and its drafting orchestration layer — the paid AI toolkit only covered the editor surface, not the differentiating logic.
Durable execution without a workflow engine: Saved ~$500/month in cloud overhead building durable execution directly on MongoDB checkpoints and Azure Service Bus instead of adopting Temporal — sustains 3-hour long-running agents with automatic crash-resume and zero data loss.
MCP for secure tool access: Standardized on Model Context Protocol (MCP) for secure tool access — deployed Dockerized servers in client VPCs with SSE tunnels for JSON-RPC schema discovery, executing agents without hardcoded tools or raw database credentials.
Blind-eval-gated architecture decisions: Gated architecture decisions on strict blind-evals — defined accuracy/cost thresholds upfront against 550 production documents, forcing a data-driven choice between ReAct and structured-output rather than rationalizing decisions post-hoc.
Semantic caching for latency and cost: Held P99 latency at 50ms via an L2 Redis Stack semantic cache. Achieved a 20-30% hit rate on repetitive queries, offsetting ~$15,000/month in LLM inference spend using a $250/month cache at million-call scale.
Structural multi-tenant isolation: Fan-out parallel search cut retrieval latency 1.3s to 443ms, then closed the multi-tenant gap structurally — session junctions are injected server-side into the vector pre-filter, so a query can narrow the search space but never widen it across tenants.
Model-tiering cost cut: Cut per-unit inference cost 5-10x via model-tiering — swapped to a cheaper model in extraction while keeping a stronger model for routing. Validated against versioned pricing across hundreds of legal Acts, scaling toward India's full statutory corpus.
Deterministic tool-execution gates: Blocked hallucinated-package execution across federated agents — implemented deterministic glass-box gates on an Agent-to-Agent (A2A) bus, running dependency-provenance checks against an allow-list before any untrusted payload executes.
Prior experience
Python Developer (contract) · AlgoHype Analytics (Jul – Aug 2024): Engineered synthetic-data generation pipelines using Azure OpenAI and shipped a QA microservice for automated data validation.
ML Engineer (short-term) · Omdena (Nov – Dec 2023): Developed an LLM-powered interview-prep chatbot, architecting the conversational flow and retrieval logic.
04 — Open-source & engineering projects
Argus — Multi-Agent Deep-Research Engine
Hit flawless context recall on a 48-item research benchmark with a planner, parallel-searcher, synthesizer orchestration engine, hand-built in async Python with no LangChain dependency.
Gates every deployment on 8 eval metrics — RAGAS faithfulness, context precision, and an LLM-as-a-judge pipeline, including abstention scoring so the system declines unanswerable queries instead of guessing.
LLM-as-a-Judge Calibration Engine
Moved judge agreement from Kappa 0.12 to 0.9261 across 5 architecture iterations — built a 10-class voice-agent intent taxonomy with blind field stripping and a stratified golden set to make the calibration honest.
Replaced manual prompt tuning with a closed loop — an offline GEPA (Generate, Evaluate, Propose, Accept) cycle reads judge failure traces and rewrites system prompts itself, compiled through a deterministic DSPy graph with enforced output typing.
Published a zero-dependency agent runtime to PyPI — bare-agent compiles any complex multi-agent graph (utilizing LangGraph/MCP patterns) into a highly durable Python orchestration script (tool registry, 3-axis budget, HITL permission gates, LiteLLM gateway).
Made the agent loop hermetically testable — a stateless reducer over an explicit message list gives free serialization for durability and eject-to-code compilation, verified by 29 tests with LLM and Redis fully faked, no daemon required.
Shipped a visual studio that compiles to real code — a companion React Flow app (Next.js 16 / React 19) for graph-based agent authoring with live SSE token streaming, plus an Eject button producing a machine-verified standalone script.
05 — Education
Bachelor of Technology in Computer Science Engineering (CSE)
Graduated 2023Parul University
Foun
ding
AI
Engi
neer
22 MO.solo build sprint
170Kdocument corpus
2,000+platform users
κ 0.9261judge calibration
27%prompt token cut
22 MO.solo build sprint
170Kdocument corpus
2,000+platform users
κ 0.9261judge calibration
27%prompt token cut
Founding
AI Engineer
22 MO.solo build sprint
170Kdocument corpus
2,000+platform users
κ 0.9261judge calibration
27%prompt token cut
22 MO.solo build sprint
170Kdocument corpus
2,000+platform users
κ 0.9261judge calibration
27%prompt token cut
01 — Relentless execution
Architecting hyperscale infrastructure in days.
Active contributions, continuous shipment and primitive-level mastery.
15,741 contributions in the last yearGitHub Contributions
01 — Work experience
Three roles. One trajectory.
Founding AI Engineer @ LawWorld
Aug 2024 – Jun 2026In-House Editor Replacement: Replaced a $12K/year vendor subscription solo-building the agentic rich-text editor backend and its drafting orchestration layer — the paid toolkit only covered the editor surface, not the differentiating logic.
Durable Agent Execution: Saved ~$500/month in cloud overhead building durable execution directly on MongoDB checkpoints and Azure Service Bus instead of adopting Temporal — sustains 3-hour long-running agents with automatic crash-resume and zero data loss.
Blind-Eval-Gated Architecture: Gated architecture decisions on strict blind-evals — defined accuracy/cost thresholds upfront against production documents, forcing a data-driven choice between ReAct and structured-output rather than rationalizing decisions post-hoc.
Semantic Caching for Latency + Cost: Held P99 latency at 50ms via an L2 Redis Stack semantic cache. Achieved a 20-30% hit rate on repetitive queries, offsetting ~$15,000/month in LLM inference spend using a $250/month cache at million-call scale.
Structural Multi-Tenant Isolation: Fan-out parallel search cut retrieval latency 1.3s to 443ms, then closed the multi-tenant gap structurally — session junctions are injected server-side into the vector pre-filter, so a query can narrow the search space but never widen it across tenants.
Model-Tiering Cost Cut: Cut per-unit inference cost 5-10x via model-tiering — routed heavy extraction to a cheaper model while keeping a stronger model for routing, validated against versioned pricing across hundreds of legal Acts.
Python Developer (Contract) @ AlgoHype
Jul 2024 – Aug 2024Synthetic-Data Generation: Engineered synthetic-data generation pipelines using Azure OpenAI, feeding downstream training and evaluation sets.
Automated QA Microservice: Shipped a QA microservice for automated data validation.
Machine Learning Engineer @ Omdena
Nov 2023 – Dec 2023LLM Interview-Prep Chatbot: Developed an LLM-powered interview-prep chatbot, architecting the conversational flow and retrieval logic.
01 — Autonomous agentic systems
Four agents that decide
their own steps.
Built solo, in production at lawworld.ai. Each runs its own loop — planning, tool calls, retrieval — and each has a runnable demo.
Lex Legal Copilot
Document-scoped retrieval restricting the agent via per-session junction tables, making cross-tenant data leakage structurally impossible.
Absolute Isolation
Zero cross-tenant leakage via junction-scoped retrieval
Lex Legal Parser
4-agent sequential pipeline choreographed over Azure Service Bus with per-page MongoDB checkpoints for crash-resilient ingestion.
722 Acts + 55 Rules Parsed
8x Speedup · 27% Token Cut · $296 Cumulative Cost
Lex Legal Drafting
4-level taxonomy routing to a bounded-ReAct generator, replacing expensive commercial editor toolkits with a custom pure-function previewer.
~$1,000/mo Displaced
Replaced commercial AI-editor toolkit · 5 Cost-Attribution Sub-stages
02 — AI backend & guided workflows
The production AI backends
underneath the agents.
FastAPI and multi-step, human-guided AI workflows — shipped solo across four backends, each feature owned end to end.
Backend AI service
Legal Search
43,927 judgments indexed
Judgment search combining keyword and semantic vector retrieval, with structured pre-filters — court, state, type, year — pushed inside the vector-search stage itself rather than applied as a slower post-filter.
hybrid keyword + vector retrievalstructured pre-filters inside $vectorSearchvoyage-4-large auto-embed
bestlaw-backend + lawworld-ai-backend
LLM eval platform
Model Researcher
Admin platform comparing multiple model-and-prompt combinations per judgment input, with full per-token trace capture and inline cost-per-comparison — no external rate-card lookup needed to know what an experiment cost.
multi-model A/B comparisonversioned prompt configper-token trace replay
bestlaw-backend
Batch AI pipeline
Case Data Generation
2,306+ Supreme Court docs
Batch-API extraction of structured case metadata at Supreme Court scale, with permanent-vs-transient failure classification so one bad document doesn't block the batch and a rate-limit doesn't get miscounted as a permanent failure.
OpenAI Batch API, 100–500 docs/batchfailure classificationpolling status scheduler
lawworld-ai-backend
Data pipeline
Human-in-the-Loop Data Factory
5-engine OCR cascade → human-verified Golden Dataset
PDF upload runs through a cascading OCR fallback chain and a YOLO v11 fine-tuned layout detector, then a human-review stage turns AI output into a Golden Dataset that feeds four downstream products — every AI-generated field carries a human-evaluated counterpart, not a confidence score standing in for one.
5-engine cascading OCR fallbackYOLO v11 fine-tuned layout detectorhuman review workflowGolden Dataset export
bestlaw-backend
01 — Open source, held to a public bar
Same primitives.
No proprietary data to hide behind.
The lawworld work above runs on synthetic demo data — real architecture, no real prompts. These three run in the open: clone the repo, run make eval or make test, and check the number yourself.
Multi-Agent Deep-Research Engine
A framework-free multi-agent deep-research engine gated by an 8-metric evaluation suite over a 48-item golden set.
1.00(Context Recall)
0.976(Faithfulness)
7/8 Gates($0 Local Ollama Run)
Python 3.12LiteLLMPostgreSQL + pgvectorFastAPI
LLM-as-a-Judge Calibration Engine
An autonomous DSPy prompt-optimization loop that iteratively closed the Cohen's Kappa gap from 0.12 to 0.9261 for voice-agent outcomes.
0.9261(Cohen's κ)
$0.149(Eval Cost / 480 Traces)
$3.10(Projected / 10K Calls vs $217)
Python 3.12DSPyLLM-as-judge
Work Experience
LawWorld
Founding AI Engineer
Architected an AI-native legal tech product end-to-end, serving 2,000+ users over a 170K-document corpus — replaced a $12K/year vendor editor in-house, built durable agent execution on MongoDB + Azure Service Bus, and cut per-unit inference cost 5-10x via model-tiering while holding P99 latency at 50ms.
Aug 2024 – Jun 2026
AlgoHype
Python Developer (Contract)
Engineered synthetic-data generation pipelines using Azure OpenAI and shipped a QA microservice for automated data validation.
Jul 2024 – Aug 2024
Omdena
Machine Learning Engineer
Developed an LLM-powered interview-prep chatbot, architecting the conversational flow and retrieval logic.
Nov 2023 – Dec 2023
Stack
Python
FastAPI
Microservices
PostgreSQL
Vector DB
PG Vector
MongoDB Atlas
Redis
Message Queue Broker
REST APIs
WebSockets
SSE
AsyncIO
OpenTelemetry
LangChain
LangGraph
Multi Agent Orchestration
MCP
DSPy
LiteLLM
AI Agents
Agentic Workflows
RAG
CRAG
Hybrid RAG
Prompt Engineering
RAGAS
LLM-as-a-Judge
Hallucination Reduction
LangFuse
Guardrails
Docker
Kubernetes
AWS
Azure
CI/CD
KEDA
System Design
Latency Tuning
LLM FinOps & Cost Optimization
Claude Code
Cursor
Git
UV
Ruff
PyTest
Next.js
React
TypeScript
Tailwind CSS
shadcn/ui
Zustand
React Query
Framer Motion
Python
FastAPI
Microservices
PostgreSQL
Vector DB
PG Vector
MongoDB Atlas
Redis
Message Queue Broker
REST APIs
WebSockets
SSE
AsyncIO
OpenTelemetry
LangChain
LangGraph
Multi Agent Orchestration
MCP
DSPy
LiteLLM
AI Agents
Agentic Workflows
RAG
CRAG
Hybrid RAG
Prompt Engineering
RAGAS
LLM-as-a-Judge
Hallucination Reduction
LangFuse
Guardrails
Docker
Kubernetes
AWS
Azure
CI/CD
KEDA
System Design
Latency Tuning
LLM FinOps & Cost Optimization
Claude Code
Cursor
Git
UV
Ruff
PyTest
Next.js
React
TypeScript
Tailwind CSS
shadcn/ui
Zustand
React Query
Framer Motion
01 — Blog
Long-form posts.
Framework-free judgment. Eval-driven design. Durable execution.
Agentic AI
Hyperscale RAG: the router is the load-bearing system, not the retriever
Hyperscale RAG systems at over a billion queries per month converge on a 4-layer meta-architecture and multi-vendor at every layer below the LLM. The substrate, the cost-cascade math, three Protocol-typed reference implementations, and the honest tradeoff against the single-vendor default.
25 min readby Subrata Mondal
03 — Contact · Status
Open to remote senior+ roles.
Applied AI · Forward Deployed Engineer · Founding Engineer
Built with
This portfolio’s tech stack
The technologies powering every animation, every snap, every pixel.
Next.js
React
TypeScript
Tailwind CSS
Motion
Bun
Vercel
Playwright
Next.js
React
TypeScript
Tailwind CSS
Motion
Bun
Vercel
Playwright
Next.js
React
TypeScript
Tailwind CSS
Motion
Bun
Vercel
Playwright
Next.js
React
TypeScript
Tailwind CSS
Motion
Bun
Vercel
Playwright
Copyright 2026 Subrata MondalBengaluru, India