Python 3.12+ (Async) · LiteLLM · MCP · DSPy Evals · Kubernetes (KEDA) · DBOS
Open to Founding / Senior / AI Engineer roles, remote, hybrid, onsite. I spent the last ~2 years as the sole AI and backend engineer at lawworld.ai, taking a legal-AI product from empty repo to production — architecture, agent orchestration, evals, CI/CD, cloud, the entire AI backend, solo — serving 2,000+ users today.
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 for 4 months, then built those 8 primitives by hand. That's what lets me name which primitive is actually failing, not just which framework.
- 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). - Cost engineering — replaced a monolithic extraction pass with a bounded Agent 0 ReAct loop, cutting total token cost 27% on a 26-document stratified eval set (26/26 exact-match retained). Per-call cost-attribution middleware into MongoDB: 100.8M+ tokens processed, $296.75 total cost, 47.6% cache hit rate.
- 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). ~2,000 registered users since the Jan 2026 launch.
- 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.
- Eval discipline — authored the eval-gating contract (ADR-005) before writing chat code: recall@5, phrase-overlap, and citation-accuracy thresholds as CI merge gates.
bare-agent — The 8-Primitive Agent Runtime
Zero-dependency agent runtime published to PyPI: strictly bounded ReAct loop, per-call cost ledgers, strict Pydantic V2 schema validation at every tool-call boundary, full transparency over the execution graph. Ships with a Next.js drag-and-drop studio that compiles visual node graphs into raw async Python.
Argus — Multi-Agent Deep-Research Engine
Permission-gated planner → parallel-searcher → synthesizer swarm (Python, LiteLLM, FastAPI, PostgreSQL/pgvector, Kubernetes + KEDA, DBOS). Cognitive routing sends extraction to fast/cheap models and reserves frontier reasoning models for synthesis, holding $0.149/turn unit economics. Clears 7 of 8 metrics on a curated eval benchmark behind a Cohen's-Kappa-calibrated LLM-judge gate. DBOS durable execution for crash-resumable research loops; KEDA scale-from-zero searcher pods on an ARQ/Redis queue.
agents-eval-framework — LLM-as-a-Judge Calibration Engine
DSPy Extractor-Judge calibration harness with an autonomous GEPA loop that reads failure traces and rewrites prompts. Drove Cohen's Kappa from 0.12 → 0.9261 across 5 architecture iterations on a 480-trace golden set.
| Domain | Technologies |
|---|---|
| Agentic & LLM Systems | Python 3.12+ (Async) · LiteLLM · MCP (Tool Calling) · DSPy · Hand-Built Agent Primitives · Hybrid RAG |
| Evaluation | LLM-as-a-Judge · Cohen's Kappa Calibration · Golden-Set Replay · RAGAS |
| Distributed & Infra | FastAPI · Docker · Kubernetes + KEDA · Azure Container Apps · Terraform · DBOS Durable Execution |
| Databases & Brokers | MongoDB Atlas · PostgreSQL (pgvector) · Redis · Kafka / Azure Service Bus |
| Observability & Security | OpenTelemetry · structlog · JWT Auth · Multi-Tenant Isolation |
role: Founding / Sole AI Engineer (lawworld.ai, Aug 2024 – Jun 2026)
status: open to Senior / Founding / AI Engineer roles, remote, hybrid, onsite
focus: Agentic Systems, Applied LLMs, Production Backends, LLM Evals, Durable Execution
timezone: IST (UTC+5:30)
contact: subratasubha2@gmail.com